JP2003050807A - Method for extracting important term/phrase/sentence - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve the problems that words and phrases not so related to the concept of a document are considered as important and the words and phrases which are just frequently used are extracted as important words and phrases since it is not always clear how much the extracted important words and phrases reflect the central concept of the document in a conventional method in the case of automatically extracting the important words and phrases from the document. SOLUTION: In this method for extracting the important words and phrases, an input document is divided into the document segments of an appropriate unit and a sum-of-squares matrix is generated from the document segments. The importance of a term and the phrase under consideration is decided on the basis of the intrinsic vector and intrinsic value of the matrix. Thus, the important term/phase/sentence related to the central concept of the document are selected.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION The present invention relates to important words, phrases, ...
The present invention relates to a method of automatically extracting a sentence, and particularly, to improve its performance by improving the expression method of a document and introducing a new measure for the importance of words and phrases.

[0002]

2. Description of the Related Art A method for automatically extracting important words or phrases from a document has been researched and developed for a long time in the field of document retrieval and information retrieval. These can be roughly divided into a heuristic method and a statistical method. As a heuristic method, there are known a method of using heading information of a document, position information of the document, and a clue expression. In the method that uses the document heading information, it is included in the title and heading based on the idea that "the document title and heading express the contents of the document concisely and contain important terms". Important terms are terms that exclude apparently insignificant terms such as articles and prepositions from the term group. This method assumes the presence of titles and headings and cannot be applied to documents that do not have them. The method of using position information in a document is to extract important words from sentences existing in the front of the article, paying attention to the fact that important sentences are intentionally written in the front of newspaper articles. To do. This method can be applied only when the location of an important part of a document is known in advance, such as a newspaper article. The method using clues is to extract such phrases by natural language analysis, assuming that a sentence starting with a particular phrase such as “As a result” is important, The scope of extraction is limited to the sentences containing them. This method cannot be applied without the underlying clues.

A statistical method that has been known for a long time is a method in which a term that frequently appears in a target document is regarded as an important word. With this method, the frequency of occurrence in documents (tf)
Is used as a measure of importance. However, this method has a problem that terms frequently appearing in one document are not always important. As a method to solve this problem, there is a tf-idf model. In the tf-idf model, the term that appears in many documents is less important, and the less the number of documents that appear is, the more important it is. For each term, find the number of documents containing that term (df), and calculate the reciprocal (id
Let f) be the term importance in the corpus, and find the product tf-idf with tf as the in-document importance to obtain the term importance. This model is a well-known method, but since it is defined by the product of the term importance in the corpus and the in-document importance, there is still the problem of how to accurately calculate the in-document importance.

[0004]

When one document is given as described above, how to obtain the in-document importance of each term becomes an important subject. The calculation of the degree of importance within a document is based on the premise that only the information included in a given document is used. The term importance in the corpus is a quantity related to the probability that each term appears in one document, and is a quantity related to the amount of information. On the other hand, since the degree of importance in a document is obtained by closing it in one document, it is desirable that it be a measure of how much the content of the document, that is, the concept of the document can be represented. Therefore, when extracting important words and phrases from a document, words and phrases representing a concept close to the concept that the document mainly represents should be preferentially extracted. For this purpose, it is essential to extract the central concept represented by the document and grasp the relationship between each word and the central concept of the document. However, in the conventional method, it is not always clear how the extracted important words reflect the central concept of the document. For this reason, words and phrases that have little relation to the concept of the document are regarded as important, and words and phrases with a high frequency are extracted as important words and phrases.

[0005]

In order to solve the above problems, the method of extracting important words according to the present invention detects a term appearing in an input document and divides the input document into document segments of appropriate units. , Generate a vector of the document segment that has the frequency of appearance of the term that appears in the document segment, calculate the eigenvectors and eigenvalues of the sum of squares matrix of the document segment vector, and use a fixed number of eigenvectors for determining important terms from all the eigenvectors. Select a term vector whose component corresponding to the term of interest has a value of 1 and is 0 otherwise, or a phrase vector whose component corresponding to the term appearing in the phrase of interest has a value of 1 and is 0 Then, the product of the square of the projected value and the corresponding eigenvalue is calculated, and the importance of the term or phrase of interest is determined based on this value.

The document segment vector is a vector whose component is a value determined based on the number of times each term appears in the document segment, that is, the frequency, and represents the concept of the document segment. The most natural unit of a document segment is a sentence. Then, the eigenvectors and eigenvalues of the sum-of-squares matrix obtained for the set of document segment vectors are obtained to develop the set of document segment vectors by eigenvectors and eigenvalues orthogonal to each other. The eigenvector has a concept in itself because it is a vector expressed by a combination of terms. Since the eigenvector is unique to the document, the concept represented by the eigenvector may be called an eigenconcept. Further, the eigenvalue can be regarded as the strength of the concept represented by the eigenvector or the energy. Therefore, eigenvectors corresponding to large eigenvalues, that is, low-order eigenvectors can be regarded as representing a central concept of a document.

A value obtained by projecting a term vector or phrase vector onto a certain eigenvector is a component that the term vector or phrase vector has in the proper concept direction corresponding to the eigenvector, and the value obtained by squaring the projection value is the above component. Represents energy. Here, the product of the value obtained by squaring the projection value and the corresponding eigenvalue, the term vector of interest, or the importance to the unique concept of interest of the phrase vector,
Select important terms and phrases based on this. Therefore, terms and phrases related to the concept of the document will be selected.

[0008]

Embodiment FIG. 1 shows a first embodiment of the present invention for extracting important terms. The method of the present invention can be implemented by running a program incorporating the present invention on a general-purpose computer. FIG. 1 is a flow chart of a computer in a state where such a program is running. A block 11 is a term detection unit, a block 12 is a morphological analysis unit, and a block 13 is a document segment division unit. A block 14 is a document segment vector creation unit, a block 15 is a sum of squares matrix calculation unit, a block 16 is an eigenvalue / eigenvector calculation unit, a block 17 is a main eigenvector selection unit, and a block 18 is an importance degree calculation unit. Also block
19 represents an important term output part. An embodiment will be described below by taking an English document as an example.

Regarding the input document, first, the term detection unit
At 11, the symbol series such as words and numbers is detected from the document. Here, words and symbol sequences are generically called terms. In the case of English sentences, it is easy to detect words because the orthography for writing words separately is established. Next, the morpheme analysis unit 12 performs a morpheme analysis such as adding a word part of speech. Next, the document segmentation is performed. The most basic unit for a document segment is a sentence. In the case of English sentences, the sentence ends with a period and a space after it, making it easy to cut out a sentence. As a method of dividing into other document segments, when one sentence consists of multiple sentences, it is divided into main clauses and subordinate clauses, and multiple sentences are grouped into document segments so that the number of terms is almost the same. There are methods, such as a method of classifying regardless of the sentence so that the number of terms included from the beginning of the document is the same.

The document segment vector creation unit 14 first determines the number of dimensions of a vector to be created from the terms appearing in the entire document and the correspondence between each dimension and each term. It is not necessary to make the components of the vector correspond to all types of terms that appear at this time, and the vector is created using only the terms determined to be nouns and verbs, for example, using the result of the part-of-speech processing. You may do it. Then, determine the type of term that appears in each document segment and its frequency, and multiply the value by the weight to determine the value of the corresponding component,
Create a document segment vector. Conventional techniques can be used to give weights.

The sum-of-squares matrix calculator 15 calculates the sum-of-squares matrix of each document segment vector. Assuming that the input document in which N terms appear is divided into M document segments, the m-th document segment vector d _m (m = 1, .., M) is (d
_When expressed by _m1 , .., d _mN ) ^T , the sum of squares matrix S = (S _ij ) can be calculated by the following equation. Here, T represents the transposition of the vector.

[0012]

[Equation 1] The eigenvalue / eigenvector calculation unit 16 calculates the eigenvalue / eigenvector of the matrix S. The k-th order eigenvectors and eigenvalues obtained are Φ _k and λ _k . Since Φ ₁ is the axis that maximizes the sum of squares of the projection values when projecting each document segment vector, it represents the concept most common to each document segment. Further, λ ₁ is the sum of squares of the projection values itself, and can be regarded as representing the strength of the concept represented by Φ ₁ or the energy. Φ ₂ is the axis that maximizes the sum of squares of the projection values under the condition that it is orthogonal to Φ ₁ .
The same applies to Φ _{3 and} after. The eigenvectors thus obtained become the basis of the subspace approximating the set of document segment vectors. If the eigenvectors are used up to the Lth order, the dimensionality of the subspace is L, and the concept of the input document is expanded by L eigenvectors having the concepts that are orthogonal to each other.

The main eigenvector selection unit 17 concretely determines the value of L. If the rank of the matrix S is R, then S eigenvectors are obtained from S, so the document is originally R
Will have unique concepts. The subspace expresses the central concept of the document by abandoning (R−L) unique concepts among these and combining L unique concepts. The subspace basis vectors are eigenvectors up to the Lth order. The ratio of the central concept to the original concept is expressed by the following equation. This formula can be used as a guide when actually determining the value of L.

[0014]

[Equation 2] The importance calculation unit 18 obtains the importance of each term for the eigenconcept corresponding to the selected eigenvector. Let us focus on the n-th term w _n , and let v _n be the document segment vector in which only the term of interest appears, the vector (term vector) in which only the n-th component has the value 1 and the others have 0. First, try to find the projection energy E of all document segment vectors to v _n . E is given by the following equation.

[0015]

[Equation 3] S _nn if the same term does not appear more than once in each document segment
Is the frequency of occurrence of the n-th term w _{n in} the document, and the conventional method that regards the frequency of occurrence (tf) in the document as the importance of w _n is E
_It is equivalent to the importance of _n . In the present invention, Equation 3 is transformed as follows.

[0016]

[Equation 4] In the derivation of number 4, Is used. Equation 4 gives E as the sum of contributions from eigenconcepts of each degree. Therefore, imp (w _n ,
_Let k) be the importance of w _{n for} the kth eigenconcept, then im
p (w _n , k) can be defined by the following equation.

[0017]

[Equation 5] Here, Φ _kn is the n-th component of Φ _k . In the definition of Expression 5, imp (w _n , k) is directly affected by the value of the frequency of occurrence of w _n . Therefore, the value normalized by S _nn in _Formula 5 can be used as the importance. In this case, the importance is as follows.

[0018]

[Equation 6] The important term output unit 19 determines and outputs the important term of the input document based on the value of each imp (w _n , k). There are two possible ways to do this. (1) In the first method, a certain number of important words in order of importance are extracted and output for each unique concept up to a constant degree L according to Equation 5 or Equation 6. The problem is how many important words are extracted for each k. For example, there is a method of determining the number of important words extracted for each k according to the value of λ _k . (2) In the second method, the sum of k = 1 to R is first taken for each imp (w _n , k) according to Equation 7 or Equation 8 to define the importance for the entire document, and then for the entire document. A certain number of terms are extracted and output in descending order of importance. As described above, the lower-order eigenconcepts are considered to be closer to the central concept of the document, so it is considered that the lower-order imp (w _n , k) should be emphasized. Imp to the importance of the term w _{n to} the document
(w _n ), ω _k is the weight for the kth eigenconcept, and

[0019]

[Equation 7] Or

[0020]

[Equation 8] Can be defined by Since ω _k should be set to a larger value for lower orders, as a way of giving, for example,

[0021]

[Equation 9] Can be used. Here, since ω _k takes a smaller value as k increases, the sum from k = 1 to L in Equation 8 may be used instead of the sum from k = 1 to R.

FIG. 2 shows a second embodiment of the present invention for extracting important phrases. The method of the present invention can be implemented by running a program incorporating the present invention on a general-purpose computer. FIG. 2 is a flowchart of the computer in a state where such a program is running. A block 11 is a term detection unit, a block 22 is a morphological analysis / syntactic analysis unit, and a block 13 is a document segment classification unit. Block 14 is a document segment vector creation unit, and block 28 is an importance calculation unit. Also block
29 indicates an important phrase output part. Of these, block 11,
13 and 14 are the same as those shown in FIG. Block 2
2 performs syntactic analysis in addition to the morphological analysis performed in block 12 of FIG. 1 to detect a phrase as an object of importance evaluation. A phrase is a combination of several terms that serves as a part of speech. An important phrase is selected from the phrases detected here.

The importance calculator 28 calculates the importance of each phrase. Let p = (p _{1, ..,} p _N ) ^T be the phrase vector of one phrase consisting of multiple terms. p is a vector in which the number of occurrences in the phrase is given to the components corresponding to the terms that make up the phrase, and 0 in the other parts. Here, the importance of the phrase imp (p)
Is defined as the sum of squares of the inner products of all document segment vectors and p. imp (p) can be described as follows.

[0024]

[Equation 12] This assumes that the document energy in the phrase vector direction × phrase energy represents the degree of importance. By the way, the number 12
In the definition of, phrase length can affect importance. Therefore, we normalized the norm squared of the phrase vector.

[0025]

[Equation 13] May be used as the degree of importance. Further, the sum of squares of the inner products of all the document segment vectors that appear commonly in Expressions 12 and 13 and p is the sum of squares matrix calculated by the sum of squares matrix calculating unit 15 in FIG.
It can be calculated as follows using S = (S _ij ).

[0026]

[Numerical equation 14] Therefore, in FIG. 2, the same process as that of the sum of squares matrix calculating unit 15 in FIG. 1 may be performed, and the importance defined by the formula 12 or the formula 13 may be obtained by using the formula 14. The important phrase output unit 29 selects and outputs a certain number of phrases in descending order of importance calculated for each phrase.

In the second embodiment, important sentences can be extracted by the following procedure. That is,
In block 13, all sentences other than the document segment used in block 14 are extracted, and in block 28, instead of the phrase vector, a sentence vector whose value of the corresponding component is the number of appearances of the term in the sentence is used. You can use it.

FIG. 3 shows a third embodiment of the present invention for extracting important phrases. The method of the present invention can be implemented by running a program incorporating the present invention on a general-purpose computer. FIG. 3 is a flow chart of a computer in a state where such a program is running. A block 11 is a term detection unit, a block 22 is a morphological analysis / syntactic analysis unit, and a block 13 is a document segment classification unit. A block 14 is a document segment vector creation unit, a block 15 is a sum of squares matrix calculation unit, a block 16 is an eigenvalue / eigenvector calculation unit, a block 17 is a main eigenvector selection unit, and a block 38 is an importance degree calculation unit. Block 39 represents an important phrase output section. Of these, block 11 and blocks 13 to 17 are the same as those shown in FIG. The block 22 is the same as that shown in FIG.

The importance calculating section 38 obtains the importance of each unique phrase for the unique concept corresponding to the selected eigenvector. As in Fig. 2, let p = (p _{1, ..,} p _N ) ^T be the phrase vector of one phrase consisting of multiple terms. The importance imp (p) defined by equation 12 is

[0030]

[Equation 15] Can be written as, and imp (p) can be regarded as the sum of the importance of the phrases for each unique concept of each degree. Therefore, here, imp (p, k) is defined as follows and its calculation is performed.

[0031]

[Equation 16] Also, in the definition of Equation 16, the importance is defined by Equation 15 imp (p)
The larger the value of, and the larger the squared norm of p, the larger the tendency. Then im given by the number 14
The value of p (p) may be normalized, and imp (p, k) may be defined as a relative value that does not depend on the value of imp (p). In this case, imp (p, k) can be defined as follows.

[0032]

[Numerical formula 17] Or, to make each imp (p, k) value and the squared norm of p independent,

[Equation 18] May be used to calculate imp (p, k). Each important phrase output section 39
The important phrase of the input document is determined and output based on the value of imp (p, k). There are two possible methods for this, as in block 19 of FIG. (1) In the first method, a certain number of important phrases are extracted and output in order of importance for each eigenconcept up to a constant degree L according to Formula 16 or Formula 17. The problem is how many important phrases are extracted for each k. For example, there is a method of determining the number of extracted important phrases for each k according to the value of λ _k . (2) In the second method, first imp (p, k)
For k = 1 to R, the importance imp (p) for the entire document is redefined, and then a certain number of phrases are extracted and output in descending order of importance for the entire document. As mentioned above, the lower-order eigenconcepts are considered to be closer to the central concept of the document, so it is thought that the lower-order imp (p, k) should be emphasized. Let ω _{k be the} weight for the unique concept of order _k ,

[0033]

[Formula 19] Can be defined by To give ω _k ,
It is possible to use the method as shown in Formula 9. Where ω _k
Since k takes a smaller value as k increases, the sum from k = 1 to L in Equation 19 may be used instead of the sum from k = 1 to R.

In the third embodiment, important sentences can be extracted by the following procedure. That is,
In block 13, all sentences other than the document segment used in block 14 are extracted, and in block 38, instead of the phrase vector, a sentence vector whose value is the number of occurrences of the term in the sentence is used as the corresponding component value. You can use it. FIG. 4 is a basic configuration diagram of the important term, important phrase, and important sentence extracting device 100 of the present invention. From the input unit 110, the user inputs a document including terms, phrases, and sentences that the user wants to extract. User operation unit 1
From 30, the unit of terms, phrases, sentences, etc. to be extracted is specified. The arithmetic unit 120 extracts important terms, important phrases, and important sentences according to the present invention. The output unit 140 outputs the extracted important terms, important phrases, and important sentences.

[0035]

[Effects of the Invention] The results obtained by extracting important words by creating a document segment vector for each sentence using 44 nouns that appear twice or more from an English document consisting of 58 sentences as terms. The results were in good agreement with human senses, and words that humans think were important were extracted as important words. As described above, according to the present invention, the words along the central concept of the document are extracted as the important words, so that the ability of extracting the important words is significantly enhanced.

[Brief description of drawings]

FIG. 1 is a diagram showing a first embodiment of the present invention.

FIG. 2 is a diagram showing a second embodiment of the present invention.

FIG. 3 is a diagram showing a third embodiment of the present invention.

FIG. 4 is a basic configuration diagram of the device of the present invention.

[Explanation of symbols]

100: Important term / phrase / sentence extraction device 110: Input section 120: arithmetic unit 130: User operation unit 140: output unit

Continued front page    F-term (reference) 5B009 QA01 QA12                 5B075 ND03 NK02 NK32 NR05 NR20                       PP25 PQ75 PR04                 5B091 AA15 CA01 EA24

Claims (24)

[Claims] 1. A method for extracting important terms from an input document including one or a plurality of document segments, comprising the following steps (a) to (f): (a) the document for each document segment Generating a document segment vector having a value related to the frequency of appearance of a term appearing in the segment, and (b) generating a square sum matrix from the document segment vector, and calculating an eigenvector and an eigenvalue of the square sum matrix. And (c) selecting a predetermined number of eigenvectors and eigenvalues from the eigenvectors and eigenvalues, and (d) for the term in the input document, only the component corresponding to the term is 1 and the other Is 0
And (e) determining the importance of the term by using the predetermined number of eigenvectors, the predetermined number of eigenvalues, and the term vector, and (f) the importance. Selecting and outputting key terms of the input document using. 2. The number of the terms is N, the number of the document segments is M, and the m-th document segment vector is
d _m = (d _m1 , .., d _mN ) ^T (m = 1, .., M), where T represents the transposition of the vector, and d _mn is the frequency of occurrence of the n-th term in the document segment. , Where the sum of squares matrix is The method of claim 1, wherein the method is calculated as: 3. The degree of importance of the term with respect to the eigenvector of each degree is determined by the product of the squared value of the inner product of the eigenvector of each degree and the term vector and the eigenvalue of each degree. The method according to 1. 4. The importance of the term with respect to the entire input document is calculated by using a weighting factor calculated from the eigenvalue of each degree, and the squared value of the dot product of the eigenvector of each degree and the term vector and the degree of each of the degrees. The method according to claim 1, wherein the weighting sum is obtained by multiplying a product of an eigenvalue and an order of the predetermined number. 5. The method according to claim 3, wherein the importance of the term is normalized by a value of a diagonal term corresponding to each term in the sum of squares matrix or a value related thereto. 6. A method for extracting an important phrase from an input document containing one or more document segments, comprising the following steps (a) to (d): (a) a term appearing in the document segment. Generating a document segment vector having a value related to the frequency of appearance as a component, and (b) for a phrase in the input document, a component corresponding to a term included in the phrase is Generating a phrase vector in which the number of occurrences of a term is given and the others are 0;
(C) determining the degree of importance of the phrase by using the sum of squares of inner products of the phrase vector and all document segment vectors, (d) selecting the important phrase of the input document using the degree of importance And output. 7. The method according to claim 6, wherein the importance of the phrase is normalized by the square of the norm of the phrase vector. 8. A method of extracting an important phrase from an input document including one or a plurality of document segments, comprising the following steps (a) to (f): (a) the document for each document segment Generating a document segment vector having a value related to the frequency of appearance of a term appearing in the segment, and (b) generating a square sum matrix from the document segment vector, and calculating an eigenvector and an eigenvalue of the square sum matrix. And (c) selecting a predetermined number of eigenvectors and eigenvalues from the eigenvectors and eigenvalues, and (d) for a phrase in the input document, a component corresponding to a term included in the phrase Generating a phrase vector in which the number of occurrences of the term in the phrase is given and the others are 0, and (e) the importance of the phrase,
Obtaining using the predetermined number of eigenvectors, the predetermined number of eigenvalues, and the phrase vector, and (f) selecting and outputting an important phrase of the input document using the importance. 9. The number of the terms is N, the number of the document segments is M, and the m-th document segment vector is
d _m = (d _m1 , .., d _mN ) ^T (m = 1, .., M), where T represents the transposition of the vector, and d _mn is the frequency of occurrence of the n-th term in the document segment. , Where the sum of squares matrix is 9. The method of claim 8 calculated as: 10. The degree of importance of the phrase with respect to the eigenvector of each degree is determined by the product of the squared value of the inner product of the eigenvector of each degree and the phrase vector and the eigenvalue of each degree. The method according to 8. 11. The degree of importance of the phrase with respect to the entire input document is calculated using the weighting factor calculated from the eigenvalues of the respective degrees, and the squared value of the inner product of the eigenvectors of the respective degrees and the phrase vector, and the degree of each of the degrees. 9. The method according to claim 8, wherein the weighting sum is calculated over the predetermined number of orders of products of eigenvalues. 12. The importance of the phrase is normalized by a squared value of the norm of the phrase vector, a sum of squared dot products of all document segment vectors and the phrase vector, or a value related thereto. , 11. 13. A method of extracting an important sentence from an input document including one or a plurality of document segments, comprising the steps (a) to (d) below: (a) a term appearing in the document segment Generating a document segment vector whose component is a value related to the frequency of appearance, and (b)
Generating a sentence vector in which, with respect to the sentence in the input document, the number of occurrences of the term in the sentence is given to the component corresponding to the term included in the sentence, and 0 otherwise.
(C) determining the importance of the sentence using the sum of squares of inner products of the sentence vector and all document segment vectors, (d) selecting the important sentence of the input document using the importance And output. 14. The degree of importance of the sentence is normalized by a square value of the norm of the sentence vector.
The method described. 15. A method for extracting an important sentence from an input document including one or a plurality of document segments, comprising the following steps (a) to (f): (a) the document for each document segment. Generating a document segment vector having a value related to the frequency of appearance of a term appearing in the segment, and (b) generating a square sum matrix from the document segment vector, and calculating an eigenvector and an eigenvalue of the square sum matrix. And (c) selecting a predetermined number of eigenvectors and eigenvalues from the eigenvectors and eigenvalues, and (d) for a sentence in the input document, a component corresponding to a term included in the sentence Generating a sentence vector in which the number of occurrences of the term in the sentence is given and 0 in the other cases, and (e) the importance of the sentence is fixed to the predetermined number. Vector, said predetermined number of eigenvalues,
And a step of obtaining using the sentence vector,
(F) A step of selecting and outputting an important sentence of the input document using the degree of importance. 16. The number of the terms is N, the number of the document segments is M, and the m-th document segment vector is d _m = (d _m1 , .., d _mN ) ^T (m = 1, .., M), where T represents the transpose of the vector and d _mn represents the value associated with the frequency of occurrence of the nth term appearing in the document segment, then the sum of squares matrix is The method of claim 15, wherein the method is calculated as: 17. The degree of importance of the sentence with respect to the eigenvector of each degree is calculated by multiplying the square value of the inner product of the eigenvector of each degree and the sentence vector by the eigenvalue of each degree. 15. The method according to 15. 18. The importance of the sentence with respect to the entire input document is calculated by using a weighting factor calculated from the eigenvalue of each degree, and the square value of the inner product of the eigenvector of each degree and the sentence vector, and the degree of each of the degrees. 16. The method according to claim 15, wherein the weighting sum is obtained by multiplying the product of an eigenvalue and the predetermined number of orders. 19. The method according to claim 17, wherein the importance of the sentence is normalized by a value of a sum of squares of inner products of all document segment vectors and a sentence vector or a value related thereto. 20. An apparatus for extracting an important phrase from an input document including one or a plurality of document segments, which comprises the following means (a) to (d): (a) a term appearing in the document segment Means for generating a document segment vector having a value related to the frequency of appearance as a component, and (b) for a phrase in the input document, a component corresponding to a term included in the phrase is Given the number of occurrences of the term,
A means for generating a phrase vector that is otherwise 0; and (c) a means for determining the importance of the phrase by using the sum of squares of inner products of the phrase vector and all document segment vectors,
(D) A means for selecting and outputting an important phrase of the input document using the importance. 21. An apparatus for extracting an important sentence from an input document including one or a plurality of document segments, comprising: (a) to (f) below; (a) the document for each document segment Means for generating a document segment vector having a value associated with the frequency of appearance of a term appearing in a segment; and (b) generating a sum of squares matrix from the document segment vectors, and calculating an eigenvector and an eigenvalue of the sum of squares matrix. Means, (c) means for selecting a predetermined number of eigenvectors and eigenvalues from the eigenvectors and eigenvalues, and (d) for a sentence in the input document, a component corresponding to a term included in the sentence. A means for generating a sentence vector in which the number of occurrences of the term in the sentence is given and 0 for the others; and (e) the importance of the sentence, the predetermined number of eigenvectors, the predetermined number. Eigenvalues, and means for determining using said sentence vector, (f) means for outputting selecting key sentences of the input document using the importance degree. 22. The number of the terms is N, the number of the document segments is M, and the m-th document segment vector is d _m = (d _m1 , .., d _mN ) ^T (m = 1, .., M), where T represents the transpose of the vector and d _mn represents the value associated with the frequency of occurrence of the nth term appearing in the document segment, then the sum of squares matrix is 22. The device of claim 21, calculated as: 23. The importance of the sentence with respect to the eigenvector of each degree is obtained by multiplying the squared value of the inner product of the eigenvector of each degree and the sentence vector by the eigenvalue of each degree. 21. The device according to 21. 24. The importance of the sentence with respect to the entire input document is calculated by using a weighting factor calculated from the eigenvalue of each degree, and the squared value of the inner product of the eigenvector of each degree and the sentence vector, and the degree of each degree. 22. The apparatus according to claim 21, wherein the apparatus obtains a weighted sum of products of eigenvalues over the predetermined number of orders.