JP2009122942A - Inter-document distance calculation device and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an inter-document distance calculation device for easily registering the new names of a person and an organization or the like and new words appearing every day by calculating concept vectors at the same time as appearance, and utilizing words or the like which are important as the characteristics of a document in inter-document distance calculation. <P>SOLUTION: This inter-document distance calculation device includes:a component frequency calculation means for calculating component frequency in an input document; a vector group calculation means for calculating a vector group by assigning vectors to the components whose vectors are registered, and assigning unregistered vectors to the components whose vectors are not registered with original expressions; and a vector inner product calculation means for calculating a vector inner product by defining the inner product of the unknown vectors as a predetermined value, and defining the inner product of the unknown vector and the vector of the components whose vectors have been registered as 0. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

The present invention relates to a method for calculating the distance between data for searching and analyzing document data.

  As a method of calculating the distance between documents, the documents are decomposed into words such as nouns and verbs, pre-registered concept vectors are associated with the words, the documents are represented as a set of concept vectors, and the average vector of the concept vectors is represented in the document It is known that the distance between documents is calculated based on the Euclidean distance (or cosine) with the concept vector of another document (for example, see Non-Patent Document 1).

  As a method for calculating the distance between vector sets such as images, it is known that the distance between the vector set A and the vector set B is calculated as follows (see, for example, Patent Document 1). First, for each vector set, a representative vector set that is a vector representing the set is determined. Next, the vector set B is vector quantized with the representative vector set of the vector set A, and the quantization error is set as an error A. The vector A is vector-quantized with the representative vector set of the vector set B, and the quantization error is defined as an error B. Here, the value with the larger error A and error B is set as the distance between the vector set A and the vector set B.

Combining the contents described in the above documents, the non-patent document 1 can express a document as a set of concept vectors, and the distance between the documents can be calculated by the distance calculation method between the vector sets in Patent document 1 subsequently.
Katsuhito Bessho, Toshiro Uchiyama, Ryoji Kataoka "Concept-based extension based on co-occurrence between words and semantic attributes" Information Processing Society of Japan, Vol.SIG-ICS 144, pp.29-34, 2006 JP 2002-366955 A

  In the above prior art, it is possible to express a word registered in advance as a vector and reflect it in calculating the distance between documents. However, new names of people, organizations, etc. appear, and new words appear daily, and it is difficult to calculate and register these words and the like at the same time as they appear. For this reason, there is a problem that words or the like that should be important as document characteristics cannot be reflected in the calculation of the distance between documents.

In the present invention, a person name, an organization name, etc. newly appear, and new words appear every day, and it is easy to calculate and register a concept vector at the same time as these words appear as a feature of the document. An object of the present invention is to provide an inter-document distance calculation apparatus that can reflect important words and the like in the calculation of the distance between documents.

According to the present invention, in a device for calculating a distance between two documents when a vector is registered in advance for a word, a compound word, or a phrase that is a component of a document, the component frequency in the input document is calculated. A vector set is obtained by assigning a vector to the constituent element frequency calculating means for calculating, assigning the vector to the constituent element in which the vector is registered, and leaving an original expression and assigning an unknown vector to the constituent element in which the vector is not registered. The vector inner product is calculated by setting the inner product between the unknown vector and the vector of the component in which the vector is registered to 0 as the inner product between the unknown vector and the vector set calculating means for calculating the vector. An inter-document distance calculation device having a vector inner product calculation means.

According to the present invention, when calculating the distance between documents, a vector that represents a word registered in advance and calculating the distance between documents are new words, synthesized words, phrases, etc. that appear daily. The effect is that unregistered words, synthesized words, phrases, etc. can be expressed by unknown vectors and reflected in the distance calculation between documents, and as a result, a more appropriate distance between documents can be calculated. Play.

  The best mode for carrying out the invention is the following embodiment.

  FIG. 1 is a block diagram showing an inter-document distance calculation apparatus 100 that is Embodiment 1 of the present invention.

  The inter-document distance calculation apparatus 100 includes a document input unit 10, a word frequency calculation unit 20, a vector set calculation unit 30, a quantization error calculation unit 40, an inter-document distance determination unit 50, and a word concept base storage unit 70. And have.

  The document input unit 10 inputs a document A that is an example of a first document and a document B that is an example of a second document.

  The word frequency calculation means 20 is a means for calculating the word frequency in the input document.

  The vector set calculation means 30 calculates the vector set by assigning the vector to the above-described component in which the vector is registered, leaving the original representation and assigning an unknown vector to the component in which the vector is not registered. It is means to do.

  The vector inner product calculation means sets the inner product of unknown vectors having the same original representation to a predetermined value, sets the inner product of unknown vectors different from each other to 0, and registers the unknown vector and the vector. The vector inner product is calculated by setting the inner product with the component vector to be zero.

  The quantization error calculating means 40 assigns each word of the word B the word that is the closest to the word A, calculates the distance between the word vectors, and calculates the calculated distance between the word vectors for all the words. Is a means for dividing the added distance by the number of words and averaging them to obtain the quantization error of the word B by the upper word A.

  The inter-document distance determining unit 50 is a unit that obtains the distance between the word A and the word B based on the first quantization error and the second quantization error. That is, of the quantization error of the second document due to the word A and the quantization error of the word A due to the second document, the larger quantification error is calculated as the distance between the word A and the second document. It is means to do.

  Next, an operation for calculating the inter-document distance between the document A and the document B in the inter-document distance calculation apparatus 100 will be described.

  FIG. 2 is a flowchart illustrating the operation of the first embodiment.

  First, in S1, document A and document B are input. Using a document corpus or the like, the concept vector corresponding to the word is calculated through dimensional compression processing such as singular value decomposition using co-occurrence information between words and semantic attributes described in Non-Patent Document 1.

  In the above embodiment, a word is used. However, instead of a word, a constituent element of a document such as a synthesized word or a phrase may be used.

  FIG. 3 is a diagram illustrating the word concept base stored in the word concept base storage unit 70 in the first embodiment.

  Here, the word vector which is the concept vector of each word is normalized to length (L2 norm) 1. Similarly, since the length of the word vector (unknown vector) of the unregistered word is normalized, it is assumed that it is “1” (α in the above description). Further, the word registered in the word concept base storage means 70 is hereinafter referred to as “registered word”. The concept base may be calculated using word-word co-occurrence information.

  Next, the morphological analysis is performed on the document A and the document B, and the word frequency is obtained.

  4 and 5 are diagrams illustrating examples of word frequencies obtained by performing morphological analysis on document A and document B. FIG.

In the document A shown in FIG. 4, the word “morning ra” has newly appeared, and is not registered in the concept base storage unit 70. That is, in FIG. 4, “unregistered” is displayed in the registration number column of the word “morning ra”. In the word frequency table included in the document B shown in FIG. 5, the words “morning ra” and “amra” are unregistered words. A set of word vectors X _A corresponding to the word set S _A = {ramen, morning ra, morning, kitakata, aim} belonging to the document A,

と記載する。

It describes.

An unknown vector is assigned to the value of the word vector corresponding to “morning ra”. Here, N _A indicates the number of words included in the document A.

In the example shown in FIG. 4, the number of words is nine. Similarly, a set of word vectors X _B corresponding to the word set S _B = {ramen, breakfast, Fukushima, Kitakata, Amler} belonging to the document B is

と記載する。

It describes.

Then, the actual distance calculation is performed. First, the word vector of the document B is quantized with the word vector set belonging to the document A, and a quantization error E _{A → B} and a quantization error E _{B → A} are obtained in S2.

Specifically, when obtaining the quantization error E _{A → B} , the word closest to the word in the document A is assigned to each word in the document B, and the distance between the word vectors is calculated. The distance between the word vectors is added to all the words, and the sum of the added distances divided by the number of words is defined as a quantization error E _{A → B} of the document B by the document A.

  That is, the quantization error is assigned to each word of the second document by assigning the word closest to the word of the first document, calculating the distance between the word vectors, and calculating the distance between the word vectors to all the words. This is the sum of words and the average of the distances divided by the number of words.

When the quantization error E _{B → A} is obtained, the word closest to the word of the document B is assigned to each word of the document A, the distance between the word vectors is calculated, and the calculated distance between the word vectors is calculated. The sum of all the words and the average obtained by dividing the added distance by the number of words is defined as a quantization error E _{B → A} of the document A by the document B.

  FIG. 6 is a diagram illustrating an example of a result of quantization of the word vector set of the document B by the word vector set of the document A.

  Since the word “morning ra” exists in the document A and the document B, the word vector of the word “morning ra” that exists in the document A and the word “morning ra” that exists in the document B The distance from the word vector “” is “0”.

On the other hand, the word “Amler” is present in document B but not in document A.
In this case, since the inner product with the word vector of any word except itself is 0 by definition, the distance between words is obtained by adding the L2 norm of each word. That is, 2. In the same manner as described above, the word vector of the document A is quantized by the word vector set belonging to the document B.

  FIG. 7 is a diagram illustrating an example of a result of quantizing the word vector of the document A by the word vector set belonging to the document B.

  In FIG. 7, the distance to “Kitakata”, which is the word closest to the word “target”, is 2.2. As described above, the distance between the word vectors having a negative inner product value is larger than the distance “2” with the unregistered word.

  Then, in S2, a quantization error is obtained by dividing the total distance by the number of words.

As described above, the quantization error E _{A → B} = 0.58 of the document B due to the document A and the quantization error E _{B → A} = 0.39 of the document A due to the document B are obtained.

  In S3, a distance is selected. That is, in S4, 0.58, which has the larger quantization error, is output as the distance between the document A and the document B.

When the quantization error E _{A → B} and E _{B → A} are expressed by equations,

である。

It is.

  FIG. 8 is a block diagram showing an inter-document distance calculation apparatus 200 that is Embodiment 2 of the present invention.

  In the first embodiment, quantization is performed using all words. The second embodiment is an embodiment in which a representative vector that represents the word vector set of each document with a smaller number of word vectors is used for the word vector set to be quantized. Thereby, the calculation cost can be reduced.

  The inter-document distance calculation device 200 is a device in which the representative vector calculation means 60 is added to the inter-document distance calculation device 100.

  The representative vector calculation means 60 selects (or calculates) a predetermined number of vectors (representative vectors) representing the distribution of the vector set obtained by the vector set calculation means 30.

  FIG. 9 is a diagram showing five representative vectors representing the words included in the document B in the inter-document distance meter device 200.

  Since the word vectors of the words “Fukushima” and “Kitakata” are similar, this is represented by one vector. When calculating a representative vector, the weighted average of each word vector is taken according to the appearance frequency of the representative word.

  Since the example shown in FIG. 9 is a schematic example, the number of representative words and the number of representative vectors are almost the same, but more ordinary words (original expressions are different) in ordinary documents. If the representative vector is calculated, for example, by clustering a set of word vectors, the number of word vectors can be greatly reduced.

However, since unregistered words are equidistant from any word vector by definition, one representative vector is assigned independently. Corresponds to “Morning Ra” and “Amler”. The representative vector of document A is denoted as {W ₁ ^A ,..., W _a ^A }, and the representative vector of document B is denoted as {W ₁ ^B ,..., W _b ^B }.

An example in which all the words of a document are representative vectors is the first embodiment. At this time, the quantization error E _{A → B} when the document B is quantized with the representative vector of the document A, and the quantization error E _{B → A} when the document A is quantized with the representative vector of the document _B Is

となる。

It becomes.

The larger one of quantization error E _{A → B} and E _{B → A} is output as the distance between document A and document B.

  Here again, it is assumed that the word vector corresponding to the unregistered word has an inner product of an arbitrary word vector excluding the same word as itself and the representative vector is zero. Using this assumption, the distance between word vectors is calculated. That is, the distance between the same words is 0, and the distance to the word vector (representative vector) of different words is the sum of the L2 norms of both vectors.

  FIG. 10 is a flowchart illustrating a flow of a series of processes in the second embodiment.

The flowchart shown in FIG. 10 is the same as the flowchart shown in FIG. 2, the process of inputting document A and document B (S 1), and the process of calculating quantization errors E _{A → B} and E _{B → A} (S 2). Between these, a process (S11) for calculating a representative vector is inserted.

  Also, with respect to the word vector set to be quantized, a representative vector for quantization error approximation that represents the word vector set of each document by a smaller number of word vectors may be used. Note that the representative vector for quantization error approximation can be made common with the representative vector for quantization.

  A representative vector for approximation of quantization error of document A

と表し、文書Ｂの量子誤差近似用代表ベクトルを

And the representative vector for approximation of quantum error in document B

と表す。

It expresses.

  For each of the documents A and B, the number of words represented by each representative vector for quantization error approximation is expressed as follows:

と表す。

It expresses.

  FIG. 11 is a diagram showing a representative vector for quantization error approximation of document B and the number of representative words.

Here, a quantization error E ₀ ^A when the word vector of the document A is quantized with the representative vector for quantization error approximation of the document A is calculated by the following equation (4).

上記と同様にして文書Ｂについても量子化誤差Ｅ_０ ^Ｂを計算する。

The quantization error E ₀ ^B is calculated for the document B in the same manner as described above.

Using the above information, the quantization error E _{A → B} when the document B is quantized with the representative vector of the document A, and the quantization error E _{B →} when the document A is quantized with the representative vector of the document _{B → A} is approximated by the following equation (5).

量子化誤差Ｅ_Ａ→ＢとＥ_Ｂ→Ａとのうちで、大きい方を、文書Ａと文書Ｂとの距離として出力する。

The larger one of quantization error E _{A → B} and E _{B → A} is output as the distance between document A and document B.

That is, in the above embodiment, the inner product of the known word x and the unknown word y is set to zero, and the distance between the unknown words (y) is set to a constant value α. The distance between the known word x and the unknown word y is (xy) ^2. This (xy) ² is expanded into x ² , 2xy, and y ² , 2xy is zero, and y ² is a constant. Assuming α, the distance between the known word x and the unknown word y is x ² + α.

  Words that are unregistered components are also represented by unknown vectors, and the inner product of the unknown vector and the vector of another component (either registered or unregistered component) is 0. The inner product of unknown vectors corresponding to the same constituent elements is set to a predetermined value α. Thereby, unregistered words can be reflected in the calculation of the distance between documents.

  If the inner product of another vector and its own vector is defined as described above, the concept of distance calculation between the vector set A and the vector set B described in Patent Document 1 is expressed as (vector set + unknown Of vector) A and (vector set + unknown vector set) B can be applied to the distance calculation. That is, the Euclidean distance D between the registered vector x and the unknown vector y is obtained by the following equation (1).

D = || xy− | ²
= || x || ² -2xy + || y || ²
= || x || ² + α Equation (1)
In other words, the concept described in Patent Document 1 is applied to distance calculation between documents.

  In other words, in the above-described embodiment, in a device that calculates a distance between two documents when a vector is registered in advance for a word, a composite word, or a phrase that is a component of the document, the configuration in the input document A component frequency calculating means for calculating an element frequency, assigning the vector to the above-described component in which a vector is registered, leaving an original expression and assigning an unknown vector to a component in which the vector is not registered A vector set calculating means for calculating a vector set and an inner product between unknown vectors as a predetermined value, and an inner product between the unknown vector and a vector of a component in which the vector is registered as 0 This is an example of an inter-document distance calculation device having a vector dot product calculation means for calculating.

  In this case, the document is the first document and the second document, and the word closest to the word of the first document is assigned to each word of the second document, and the distance between the word vectors is calculated. Then, the calculated distance between the word vectors is added up for all the words, and the average of the added distance divided by the number of words is calculated as the first quantization error. The word closest to the word in the second document is assigned to the word, the distance between the word vectors is calculated, the calculated distance between the word vectors is added for all the words, and the added distance is calculated. There is provided a quantization error calculating means for calculating the average obtained by dividing by the number of words as the second quantization error.

  An inter-document distance determining unit is provided that determines a large quantization error among the first quantization error and the second quantization error as a distance between the first document and the second document. It has been.

  Further, in the above embodiment, in a device that calculates a distance between two documents when a vector is registered in advance for a word, a composite word, or a phrase that is a component of the document, the configuration in the input document A component frequency calculating means for calculating an element frequency, assigning the vector to the above-described component in which a vector is registered, leaving an original expression and assigning an unknown vector to a component in which the vector is not registered The vector set calculation means for calculating the vector set is the same as the original expression, the inner product of the unknown vectors having the same original expression is set to a predetermined value, the inner product of the unknown vectors having different original expressions is set to 0, and the unknown vector In the vector to calculate the vector inner product with the inner product of the vector and the component vector in which the vector is registered as 0 It is an example of a document between the distance calculation unit and a calculation unit.

  In this case, the document is the first document and the second document, and the word closest to the word of the first document is assigned to each word of the second document, and the distance between the word vectors is calculated. Then, the calculated distance between the word vectors is added up for all the words, and the average of the added distance divided by the number of words is calculated as the first quantization error. The word closest to the word in the second document is assigned to the word, the distance between the word vectors is calculated, the calculated distance between the word vectors is added for all the words, and the added distance is calculated. There is provided a quantization error calculating means for calculating the average by dividing by the number of words as the second quantization error.

  In addition, there is an inter-document distance determining unit that determines that a large quantization error among the first quantization error and the second quantization error is a distance between the first document and the second document. .

  Further, in the above embodiment, in the apparatus for calculating the distance between two documents when a vector is registered in advance for a word, compound word, or phrase that is a component of the document, the vector set of the first document A representative vector set calculating means for obtaining a first representative vector set for quantization and calculating a second representative vector set for quantization from the vector set of the second document; A component frequency calculating means for calculating a component frequency, and assigning the vector to the above-described component in which the vector is registered, and assigning an unknown vector while leaving the original expression for the component in which the vector is not registered Thus, the inner product of the unknown vectors having the same original representation as the vector set calculating means for calculating the vector set is determined in advance. A vector dot product calculating means for calculating a vector dot product by setting the dot product of the unknown vectors whose values are different from each other to 0 as the value, and setting the dot product of the unknown vector and the vector of the component in which the vector is registered as 0 It is an example of the distance calculation apparatus between documents which has.

  In this case, a second quantization error when the vector set of the first document is quantized by the second representative vector set for quantization is obtained, and the first quantization vector set is used. A first quantization error when the vector set of the second document is quantized is obtained, and a first representative vector set for the quantization is used to quantize the vector set of the second document. Quantization error calculation means is provided for obtaining a second quantization error when a first quantization error is obtained and the vector set for the second document is quantized by the second quantization vector set.

  Further, the inter-document distance calculation means calculates a larger quantification error among the quantization error of the second document by the first document and the quantization error of the first document by the second document. A means for calculating a distance between the document and the second document.

  In the above embodiment, the representative vector set for quantizing the document is quantized from the vector set of the first document in the apparatus for calculating the quantization error of the vector set of another document at high speed. The first representative vector set of the second representative vector set is obtained by obtaining the frequency number of words represented by each representative vector and represented by the second representative vector set to be quantized and each representative vector from the vector set of the second document. A means for obtaining a frequency number of words to be quantized, a first representative vector set to be quantized, a quantization error used for the first approximation when the first vector set is quantized, A second representative vector set for obtaining a quantization error used for the second approximation when the second vector set is quantized, a quantization error used for the first approximation, and the quantum The first to be The representative vector set of examples of fast quantization error computing device having means for obtaining an approximate quantization error when quantizing the second representative vector set for quantizing.

Note that the quantization error used in the first approximation is a quantization error E ₀ ^A shown in Expression (4). The quantization error used for the second approximation is a quantization error E ₀ ^B calculated in the same manner as the quantization error E ₀ ^A. The approximate quantization error is the quantization error E _{A →} B and E _{B → A} shown in Expression (5).

Moreover, the said Example can be grasped | ascertained as invention of a program. In other words, the above embodiment is an example of a computer program for causing a computer to execute functions as each means constituting the inter-document distance calculation apparatus. Furthermore, the said Example is an example of the computer program for making a computer perform the function as each means which comprises the said quantization error calculation apparatus.

1 is a block diagram illustrating an inter-document distance calculation apparatus 100 that is Embodiment 1 of the present invention. FIG. 3 is a flowchart showing the operation of the first embodiment. It is a figure which shows the word concept base stored in the word concept base storage means 70 in Example 1. FIG. FIG. 6 is a diagram illustrating an example of word frequencies obtained by performing morphological analysis on a document A and a document B; FIG. 6 is a diagram illustrating an example of word frequencies obtained by performing morphological analysis on a document A and a document B; FIG. 10 is a diagram illustrating an example of a result of quantization of a word vector set of document B by a word vector set of document A. 10 is a diagram illustrating an example of a result of quantizing a word vector of a document A by a word vector set belonging to the document B. FIG. It is a block diagram which shows the distance calculation apparatus 200 between documents which is Example 2 of this invention. 5 is a diagram showing five representative vectors representing words included in a document B in the inter-document distance meter device 200. FIG. 10 is a flowchart illustrating a flow of a series of processes in the second embodiment. FIG. 6 is a diagram illustrating a representative vector for quantization error approximation of a document B and a representative number of words.

Explanation of symbols

100: Inter-document distance calculation device,
10 ... Document input means,
20 ... Word frequency calculation means,
30: Quantization error means,
40: Quantization error calculation means,
50 ... Document distance determination means,
60 ... representative vector calculating means,
70: Word concept base storage means,
200: Inter-document distance calculation device.

Claims (9)

In a device for calculating a distance between two documents when a vector is registered in advance for a word, a compound word, or a phrase that is a component of the document,
Component frequency calculation means for calculating the component frequency in the input document;
Vector set calculating means for calculating a vector set by assigning the vector to the component in which the vector is registered, and leaving an original expression and assigning an unknown vector to the component in which the vector is not registered;
The inner product of the unknown vectors having the same original representation is set to a predetermined value, the inner product of the unknown vectors having different original representations is set to 0, and the vector of the component in which the unknown vector and the vector are registered A vector dot product calculating means for calculating a vector dot product with an inner product of 0 as a vector;
An inter-document distance calculation apparatus comprising: In claim 1,
The document is a first document and a second document,
Assigning to each word in the second document the word closest to the word in the first document, calculating the distance between the word vectors, and adding the calculated distance between the word vectors for all words; The average distance obtained by dividing the sum of the distances by the number of words is calculated as the first quantization error,
Assigning to each word in the first document the word closest to the word in the second document, calculating the distance between the word vectors, and adding the calculated distance between the word vectors for all words; An inter-document distance calculation device comprising quantization error calculation means for calculating an average obtained by dividing the added distance by the number of words as a second quantization error. In claim 2,
An inter-document distance determining means for determining a large quantization error among the first quantization error and the second quantization error as a distance between the first document and the second document; Characteristic distance calculation device. In a device for calculating a distance between two documents when a vector is registered in advance for a word, a compound word, or a phrase that is a component of the document,
Representative vector set calculating means for obtaining a first representative vector set for quantization from the vector set of the first document, and calculating a second representative vector set for quantization from the vector set of the second document; ;
Component frequency calculation means for calculating the component frequency in the input document;
A vector set calculating means for calculating a vector set by assigning the vector to the above-mentioned component in which the vector is registered, and assigning an unknown vector to the component in which the vector is not registered, leaving an original expression; ;
The inner product of the unknown vectors having the same original representation is set to a predetermined value, the inner product of the unknown vectors having different original representations is set to 0, and the vector of the component in which the unknown vector and the vector are registered A vector dot product calculating means for calculating a vector dot product with an inner product of 0 as a vector;
An inter-document distance calculation apparatus comprising: In claim 4,
A second quantization error when the vector set of the first document is quantized is obtained by the second representative vector set for quantization, and the first quantization vector set is used to obtain the second quantization error. Find the first quantization error when quantizing a vector set of two documents,
A first quantization error is obtained when the vector set of the second document is quantized by the first representative vector set for quantization, and the second quantization vector set is used for obtaining the first quantization error. An inter-document distance calculation apparatus comprising quantization error calculation means for obtaining a second quantization error when a vector set of one document is quantized. In claim 4,
The inter-document distance calculation means calculates a larger quantification error between the quantization error of the second document by the first document and the quantization error of the first document by the second document, and the first document. An inter-document distance calculation apparatus, which is a means for calculating a distance from the second document. In a representative vector set for quantizing a document and calculating a quantization error of a vector set of another document at high speed,
The first representative vector set to be quantized is obtained from the vector set of the first document, the frequency number of words represented by each representative vector is obtained, and the quantized from the vector set of the second document Means for determining a second representative vector set and the frequency of words represented by each representative vector;
In the first representative vector set to be quantized, the quantization error used for the first approximation when the first vector set is quantized, and the second representative vector set to be quantized, Means for determining a quantization error used for the second approximation when the second vector set is quantized;
An approximate quantization error when the quantization error used for the first approximation and the first representative vector set to be quantized are quantized by the second representative vector set for quantization is expressed as follows. Means to seek;
A high-speed quantization error calculation device characterized by comprising:   The computer program for making a computer perform the function as each means which comprises the distance calculation apparatus between documents of any one of Claims 1-6.   A computer program for causing a computer to execute a function as each means constituting the quantization error calculation device according to claim 7.

Images