JP2007133516A - Document sorting method, document sorting program and document sorting device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a document sorting method for reducing any influence due to the variations of the notation of a document. <P>SOLUTION: Similarity based on a common keyword between two documents among documents to be sorted and similarity based on a common cited document are calculated. The mean value of those similarity is defined as document similarity between the two documents. An stable inter-document distance is calculated based on the document similarity for each of those two documents. The documents to be sorted are arranged in a coordinate system, and each document is moved so that each of the two documents has the stable inter-document distance, thereby enabing the strongly relevant documents to be gathered. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a document classification method for classifying a plurality of documents to be classified according to their contents.

Conventionally, there has been a document classification method for calculating similarity between documents based on an evaluation value of non-generality of a common keyword.
Japanese Patent Laid-Open No. 8-222447 Japanese Patent No. 3385297 JP 2003-288352 A JP 2002-312803 A

  However, according to the conventional document classification method, when the contents of the documents are similar (for example, in the case of patent documents in which the subgroup of the international patent classification is the same), the influence of the fluctuation of the notation becomes large and appropriate document classification cannot be performed. There was a problem.

  The present invention has been made in view of the above problems, and an object of the present invention is to provide a document classification method that is less affected by fluctuations in document notation.

  The document classification method of the present invention is a document classification method for classifying a plurality of classification target documents according to their contents, and extracts a common keyword that appears in common among two of the classification target documents. An extraction step, a common cited reference extracting step for extracting common cited references that are commonly cited in two of the documents to be classified, and a word included in each keyword reference document with an identification number of the keyword reference document The number of occurrences of the common keyword in the keyword table is counted with reference to the keyword table stored in association with the keyword table, and the non-generality evaluation value of the common keyword is evaluated based on the number of occurrences (evaluation as the number of occurrences increases). Value = weight is set to be small)) and a common keyword evaluation value calculating step for calculating Referring to the cited document table that stores the documents cited in the document reference document in association with the identification number of the cited document reference document, count the number of appearances of the common cited document in the cited document table, A common citation evaluation value calculation step for calculating an evaluation value of non-generality of the common citation based on the evaluation value (the evaluation value is set so that the weight becomes smaller as the number of appearances increases); A keyword similarity calculation step of calculating the similarity based on a common keyword of two documents of the classification target documents by adding the evaluation values of the keywords; and adding the evaluation values of each of the common cited documents, A cited document similarity calculating step for calculating a similarity based on a common cited document of two documents of the target document, a similarity based on a common keyword, and a common quote A document similarity calculation step for calculating a document similarity that combines the similarity based on the contribution, and the classification target document is arranged in a coordinate system, and the two documents having a high document similarity approach each other, and the document similarity is low A document moving step of performing a process of moving each of the classification target documents so that the two documents move away from each other. By combining the similarity based on the common keyword and the similarity based on the common cited document, it is possible to reduce the influence of the fluctuation of the notation. Further, since the similarity based on the common keyword is also reflected in the document similarity, an appropriate similarity can be acquired even for documents with different target fields.

  In the present invention, prior to combining the similarity based on the common keyword and the similarity based on the common cited document, it is preferable to reduce the weight by dividing the similarity based on the common keyword by an appropriate number. Thereby, after performing rough document classification based on the similarity based on the common keyword, the same effect as performing fine document classification based on the similarity based on the common cited document can be obtained.

  The preferred form of the present invention is a form obtained by adding any of the constituent elements in the dependent claims (any combination of constituent elements in the dependent claims) to the form specified by the independent claims in the claims. Including.

  It is possible to obtain an appropriate degree of similarity even for documents with different target fields, as well as the influence of fluctuations in document notation is reduced.

  Hereinafter, preferred embodiments of the present invention will be described in detail with reference to the accompanying drawings. FIG. 1 is a block diagram showing an embodiment of a document classification apparatus according to the present invention. The document classification device 1 classifies a plurality of classification target documents according to the contents of each classification target document. The document classification device 1 includes a database 10, a stable inter-document distance calculation unit 22, an arrangement document selection unit 23, a position coordinate initial value setting unit 24, an inter-document force vector calculation unit 26, and a position coordinate update unit 28. The database 10 includes a classification target document DB 12, a stable inter-document distance DB 14, a position coordinate DB 16, and an inter-document force vector DB 18. The classification target document DB 12 stores a plurality of classification target documents in association with document codes that specify each document. The classification target document is stored in the classification target document DB 12 in advance, but can be input as necessary by an appropriate input unit.

  The stable inter-document distance DB 14 stores the stable inter-document distance calculated by the stable inter-document distance calculation unit 22 in association with the document code. FIG. 2 is a configuration diagram illustrating an example of the database of the inter-document distance DB 14. As shown in this figure, the stable inter-document distance between two documents is stored in association with the document codes (P0001, P0002,...). For example, in this case, the stable inter-document distance between the document (P0001) and the document (P0002) is 0.005.

  The position coordinate DB 16 stores the initial value of the position coordinate of each document set by the position coordinate initial value setting unit 24 and the position coordinate updated by the position coordinate update unit 28 in association with the document code. FIG. 3 is a configuration diagram illustrating an example of the database of the position coordinate DB 16. As shown in this figure, the position coordinates (X coordinate, Y coordinate) of each document are stored in association with the document code. For example, in this case, the position coordinates of the document (P0003) are (0.5155, 0.3417).

  The inter-document force vector DB 18 stores the total inter-document force vector calculated by the inter-document force vector calculation unit 26 in association with the document code. FIG. 4 is a configuration diagram illustrating an example of the database of the total inter-document force vector DB 18. As shown in this figure, the total inter-document force vector (FX, FY) working on each document is stored in association with the document code. For example, in this case, the total inter-document force vector of the document (P0002) is (0.007, -0.003).

  The stable inter-document distance calculation unit 22 calculates a stable inter-document distance between two documents for a plurality of classification target documents stored in the classification target document DB 12 according to the degree of similarity between the two documents. The stable inter-document distance decreases as the degree of similarity between both documents increases and decreases as the degree of similarity decreases.

The position coordinate initial value setting unit 24 sets the initial value of the position coordinate of each document on the two-dimensional coordinate plane. An example of an initial value setting method in the position coordinate initial value setting unit 24 will be described. For convenience of explanation, the number of documents to be classified is N (N is an integer of 2 or more), and each document is represented as T _i (i = 1, 2,..., N). First, the document _{T i} and the document _T j reads (j = 1,2, ···, N , j ≠ i) stable document distance _L 0 (i, j) between the in table _{L a.} After reading the stable inter-document distance L ₀ (i, j) for all (i, j) pairs, the average value L _avg of L ₀ (i, j) is obtained. Then, the position coordinates (X _i , Y _i ) of each document T _i are expressed by the following equation:
X _i = L _avg × rnd
Y _i = L _avg × rnd
Ask from. Here, rnd represents a random number. Thereby, the initial value of the position coordinate of each document is set. The stable inter-document distance L ₀ (i, j) is normalized by dividing by the average value L _avg .

  The inter-document force vector calculation unit 26 calculates a total inter-document force vector that works on each document. The total inter-document force vector is a vector sum of inter-document forces that each document receives from other documents. Further, the inter-document force means that when the distance on the coordinate plane obtained from the position coordinates of each of the two documents is larger than the above-mentioned stable inter-document distance, an attractive force acts between both documents, and conversely, the distance on the coordinate plane This is a concept introduced on the assumption that repulsive force acts between both documents when the distance is smaller than the distance between stable documents. The magnitude of these forces increases as the absolute value of the difference between the distance on the coordinate plane and the stable inter-document distance increases and decreases as the absolute value decreases. When the distance on the coordinate plane matches the stable inter-document distance, the inter-document force acting between the two documents is zero.

An example of a method for calculating the inter-document force vector in the inter-document force vector calculation unit 26 will be described. First, the distance L (i, j) between the document T _i and the document T _j is referred to as a processing time point (current processing time point) (in particular, in the present embodiment, the update of position coordinates is referred to as “current processing time point” and “next processing time point”). The “current processing time point” refers to the movement process (the process of calculating the total inter-document force vector and updating the position coordinates based on the total or part of each of the arranged documents is determined as the convergence of the position coordinates. In the repetitive processing of (repeating until a certain time), the time point at which a certain time starts is referred to, and the “next processing time point” indicates the time point at which the next time after the certain time starts). Based on the coordinates,
L (i, j) = {(X _i −Y _i ) ² + (X _j −Y _j ) ² } ^0.5
Ask from. Incidentally, the term "based on the position coordinates of the two in the processing time," as will be described later, the position coordinates are updated as required in each document T _i, always take the same value Because there is no. Next, the inter-document force f (i, j) between the document T _i and the document T _j is expressed by
f (i, j) = (L ₀ (i, j) −L (i, j)) / (L ₀ (i, j) + ε1) ^α
Ask from. Here, ε1 is a constant to cope with when L ₀ (i, j) is 0, and is set to 1 × 10 ⁻¹² , for example. α is set so that the inter-document force f (i, j) increases exponentially as the stable inter-document distance L ₀ (i, j) decreases. By doing so, a greater inter-document force works when the similarity between documents is high. As a result, it becomes easy to form a group of similar documents, the position where the group is arranged is close to the human sense, and it can be easily converged even if the number of classified documents N increases. it can. When the number of classification target documents is relatively small (when N is less than 50), α is set to any value between 0.8 and 2.3. When N exceeds 100, it can be easily converged by setting α to any value of 1.8 to 2.2. In particular, when N = 101 to 3000, α = 2 is preferable. In particular, when mapping a document to be classified into a two-dimensional space, if α is smaller than the above range, there are many cases where the coordinates of some documents do not converge and diverge in the course of repeated movement processing. If it is larger, it becomes easier to form a uniform document group that does not reflect the contents of individual documents. Next, the X component fX (i, j) and the Y component fY (i, j) of the inter-document force that the document T _i receives from the document T _j are expressed by the following equations:
fX (i, j) = f (i, j) × (X i -X j) / (L (i, j) + ε2) β
fY (i, j) = f (i, j) × (Y i -Y j) / (L (i, j) + ε2) β
Ask from. Here, ε2 is a constant to cope with L (i, j) being 0, for example, 1 × 10 ⁻¹² . Β is set to 0.5, for example. Finally, the X component FX _i and the Y component FY _i of the sum of the inter-document forces acting on each document T _i are expressed by the following equations:
FX _i = Σ _j fX (i, j)
FY _i = Σ _j fY (i, j)
Ask from. Here, Σ _j means taking the sum of all placed documents. The vector having FX _i and FY _i calculated in this way as components is the above-described total inter-document force vector.

The position coordinate update unit 28 updates the position coordinates of each document so that the absolute value of the total inter-document force vector calculated by the inter-document force vector calculation unit 26 becomes small. An example of the position coordinate update method in the position coordinate update unit 28 will be described. That is, the position coordinates (X _i , Y _i ) of each document T _i are expressed by the following equation based on the inter-document force vector (FX _i , FY _i ) calculated by the inter-document force vector calculation unit 26:
X _i ′ = X _i −k × FX _i
Y _i ′ = Y _i −k × FY _i
Updated by Here, (X _i ′, Y _i ′) represents the updated position coordinates. Further, k is a movement coefficient, for example, a constant not less than 1 × 10 ^{−23 and not} more than 1 × 10 ⁻²² . The above formula, each document T _i, in the direction of the document between the force vector, which means that is moved by a distance proportional to the magnitude of the absolute value of the vector. The updated position coordinates are stored in the position coordinate DB 16 and are overwritten on the position coordinates stored so far. Position coordinate update section 28 in the present embodiment, in conjunction with the updated coordinates, an average value ML of the movement distance of each document T _i the following formula,
ML = Σ {(k × FX _i ) ² + (k × FY _i ) ² } ^0.5
Ask from. This average value ML is used when determining the convergence condition by the convergence condition determination unit 30 described later.

  The document classification device 1 further includes a convergence condition determination unit 30, a display unit 32 (output unit), and an input unit 34. The convergence condition determination unit 30 determines the convergence condition after the position coordinates are updated by the position coordinate update unit 28. For example, it can be set as a convergence condition that the average value ML obtained by the position coordinate update unit 28 is equal to or less than a specified value. When the convergence condition is not satisfied, the convergence condition determination unit 30 causes the inter-document force vector calculation unit 26 to calculate the total inter-document force vector again using the updated position coordinates, and causes the position coordinate update unit 28 to The position coordinates are updated again using the total inter-document force vector. Therefore, the update of the position coordinates by the position coordinate update unit 28 is executed until the above convergence condition is satisfied.

Display unit 32, the convergence condition described above is satisfied, after the updating of the position coordinates by the position coordinate update section 28 has been completed, based on the determined position coordinates, relative position on the coordinate plane between the document T _i Visualize and display relationships. An example of a display method on the display unit 32 will be described. FIG. 5 is a diagram illustrating an example of a result display screen by the display unit 32. In this example, first, the display area 50 is divided into m × n cells (here, m = n = 4). Further, the maximum value (X _max , Y _max ) and the minimum value (X _min , Y _min ) of the X coordinate and Y coordinate that define the display area are input by the input unit 34 described later. In addition, without inputting these values, the maximum and minimum values of the X coordinate and the Y coordinate can be used as default values from the position coordinates of all documents that have already been determined. Next, the display part 32 calculates | requires the coordinate range corresponded to each cell from these input values. Then, the number of documents included in each cell is displayed as shown in FIG. For example, in this case, the number of documents included in the upper right cell is one. Further, in this example, an image of a document included in each cell is created, and the image is hyperlinked to each cell. As shown in FIG. 5, when the mouse pointer 52 is moved to a cell of interest, the documents included in the cell can be displayed as a corresponding document list. Here, patent documents such as published patent gazettes (including patent documents issued in foreign languages or foreign countries) are assumed as classification target documents, and the type and gazette number of patent documents are displayed in the corresponding document list. I am letting. In addition, since hyperlinks are pasted on these displays, for example, if a portion displayed as “JP-A-8-XXXXX” is clicked on the screen, the published patent gazette You can access the image and see its contents.

The input unit 34 is for inputting a display area or the like on a coordinate plane to be displayed by the display unit 32, and for example, a keyboard or a mouse is used. For example, in the example of FIG. 5, values of X _max , Y _max , X _min , and Y _min that define the display area 50 can be input from the input unit 34. The input information is passed to the display unit 32.

Next, the operation of the document classification apparatus 1 will be described, and an embodiment of the document classification method according to the present invention will be described. FIG. 6 is a flowchart showing the initial process and the process of arranging and moving the initially arranged document in the two-dimensional display coordinate system. First, the stable inter-document distance calculation unit 22 reads the classification target document stored in the classification target document DB 12, calculates the stable inter-document distance between the two documents, and calculates the calculated stable inter-document distance as the stable inter-document distance. The data is stored in the DB 14 (S61). Subsequently, the stable inter-document distance calculation unit 22 reads the stable inter-document distance stored in the inter-document distance DB 14 to calculate an average value (S62), and divides each stable inter-document distance by this average value. Is normalized and the data in the stable inter-document distance DB 14 is updated (S63). The arrangement document selection unit 23 selects int√N (a value obtained by rounding down the decimal point of the square root of the total number N of the classification target documents) as the initial arrangement document T _k that is the classification target document to be first arranged in the display coordinate system ( S64). The position coordinate initial value setting unit 24 sets the initial value of the position coordinate of each document using the average value, and stores the initial value of the set position coordinate in the position coordinate DB 16 (S65). Then, the inter-document force vector calculation unit 26 reads the stable inter-document distance stored in the inter-document distance DB 14 and the position coordinates stored in the position coordinate DB 16, and uses these values to calculate the total sum applied to each document. The inter-document force vector is calculated, and the calculated total inter-document force vector is stored in the inter-document force vector DB 18 (S66). Thereafter, the position coordinate update unit 28 reads the total document inter-force vector stored in the inter-document force vector DB 18, updates the position coordinates of each document based on the vector, and stores the updated position coordinates in the position coordinate DB 16. Store (S67). When the position coordinates are updated, the convergence condition determination unit 30 determines the convergence condition. If the convergence condition is not satisfied, the above steps (S66 to S67) are repeatedly executed. If the convergence condition is satisfied, the process proceeds to a process of adding a new classification target document.

  The document classification device 1 includes a display unit 32. Thereby, the user can easily know the relative positional relationship between documents by viewing the display on the display unit 32. The document classification device 1 may be configured not to include the display unit 32. In this case, for example, an output unit that outputs the classification result may be provided instead of the display unit 32, and the output content may be displayed on an external display or printed by an external printer.

  In addition, the document classification device 1 includes an input unit 34 that can input the maximum value and the minimum value of the X coordinate and the Y coordinate that define the range displayed in the display area 50 (see FIG. 5). As a result, the user can display a desired range on the coordinate plane and know in detail the positional relationship between documents in the range.

  Note that the position coordinate update unit 28 preferably executes the update of the position coordinates until the absolute value of the inter-document force vector acting on each document becomes the minimum value obtained by summing all the documents to be classified. . In this case, the position coordinates of each document can be determined while maintaining particularly high consistency among all the classification target documents.

  FIG. 7 is a block diagram showing an example of the configuration of the stable inter-document distance calculation unit 22 in FIG. The stable inter-document distance calculation unit 22 includes a word / cited document extracting unit 70 that extracts words and cited documents from various documents, and various databases 80 that store the words and cited documents extracted by the word / cited document extracting unit 70. It has.

  The word / cited document extracting unit 70 extracts a word / cited document from the key document as a keyword / cited document and a reference / extracting the word and the cited document as a reference word / cited document from the reference document. It has a word / cited document extracting unit 72 and a search word / cited document extracting unit 73 that extracts words and cited documents from the search document as search words / cited documents. Here, the classification of “key document” and “search document” is convenient, and in the stable inter-document distance calculation unit 22, one of the two documents whose inter-document distance is desired to be obtained is the key document, and the other is Search document. The reference document includes a keyword evaluation value (a value indicating a degree that each keyword is uniquely included in the key document (non-generality)) and a cited document evaluation value (a degree to which each cited document is uniquely included in the key document ( This is a document that is referred to when setting a value representing non-generality. As the reference document, for example, all the documents in the classification target document DB 12 (see FIG. 1) or a part of the documents in the classification target document DB 12 extracted in advance at random can be used. The reference document can be input to the stable inter-document distance calculation unit 22 as necessary by an appropriate input unit. The stable inter-document distance calculation unit 22 includes storage means for storing a reference document.

  In Japanese, the extraction units 71 to 73 have a function of extracting words in a document using hiragana, punctuation marks, special symbols, and spaces as delimiters or using a morphological analysis tool or the like. The extraction units 71 to 73 extract a word in a document using a special symbol and / or space as a delimiter or a morphological analysis tool or the like in a language in which alphabets such as English are used. Have The extraction units 71 to 73 have a function of extracting only words that do not match by collating words extracted from a document with words already extracted from the same document so as not to extract duplicate words from one document. Have

  The extraction units 71 to 73 include character strings indicating the types of patent documents such as “JP,” “Patent,” “US,” “USP,” “EP,” “DE,” “JP,” “WO,” and the like. It has a function of retrieving a cited document by searching for a combination with a year display, a number, and the like subsequent thereto. In addition, when a tag indicating a cited document is defined as in recent electronic publications, a cited document can be extracted using this tag. Furthermore, it is possible to obtain a cited document for each publication using various existing databases.

  The database (DB) 80 includes a keyword DB 81, an all word DB 82, a keyword evaluation value DB 83, a search word DB 84, a keyword similarity DB 85, a cited document DB 86, a cited document evaluation value DB 87, a cited document similarity DB 88, and a document similarity DB 89. is doing. The keyword DB 81 stores keywords extracted from the key document. The keyword is stored in association with the key document code that identifies the key document that is the extraction source. The all word DB 82 (a table in which the words included in each document are stored in association with the document number of the document) includes a reference word extracted from the reference document, a reference document code for identifying the reference document from which each extraction is made, and Store in association. The keyword evaluation value DB 83 stores evaluation values calculated by a keyword evaluation value calculation unit 91 described later. The search word DB 84 stores search words extracted from the search document. The search word is stored in association with a search document code that specifies a search document as an extraction source. The keyword similarity DB 85 stores the keyword similarity calculated by the keyword similarity calculation unit 92 described later. In the cited document DB 86, a cited document code for identifying a document cited as a related document in each reference document is stored in association with the reference document code. The cited document stored in the cited document DB 86 may be extracted by the word / cited document extracting unit 70 or may be based on input data. The cited document evaluation value DB 87 stores an evaluation value calculated by a cited document evaluation value calculation unit 93 described later. The cited document similarity DB 88 stores the cited document similarity calculated by the cited document similarity calculating unit 94 described later. The document similarity DB 89 stores a document similarity that is a similarity obtained by combining the keyword similarity and the cited document similarity.

  Note that the keyword, reference word, and search word may be extracted from the entire document to be extracted, or may be extracted from a part thereof. For example, if the document to be extracted is a patent document, the extraction range may be limited to bibliographic items, summaries, claims, or examples. In particular, when the data amount is limited, it is effective to narrow the extraction range to a part of the document. In addition, the reference word is extracted from a part of the reference document, the keyword and the search word are extracted from the entire key document and the search document, respectively. Can be adjusted. When extracting cited references, it is necessary to extract bibliographic items and search reports such as the specification part excluding the examples and claims, or [Background Art], US Patent Gazette and European Patent Gazette (progress information). In the case where the part is listed as a list, it is preferable to limit to this part.

The stable inter-document distance calculation unit 22 includes a keyword evaluation value calculation unit 91 and a keyword similarity calculation unit 92. The keyword evaluation value calculation unit 91 has a function of calculating an appearance rate at which a word common to the key document appears in all reference documents. When there are N reference documents and B has a common word, the keyword appearance rate in all documents is calculated as B / N. The keyword evaluation value calculation unit 91 searches the reference words stored in the all-word DB 82 and calculates how many of the same keywords exist. The keyword appearance rate in all documents is calculated by dividing the counted number of reference words by the number of all reference documents. Further, the keyword evaluation value calculation unit 91 has a function of calculating a keyword evaluation value by taking the reciprocal of the keyword appearance rate in all documents. That is, the keyword evaluation value is calculated by N / B and indicates the degree (non-genality) in which each keyword is inherently included in the key document. It may be possible to adjust the keyword appearance number B to the power of 0.1 to 0.8. In this case, for example, when an adjustment of 0.5 power is performed, the keyword evaluation value is N / (B ^0.5 ).

  The keyword similarity calculation unit 92 has a function of calculating a keyword similarity representing the degree of similarity between the key document and the search document by adding the evaluation values of all the keywords included in the search document. Further, it is conceivable that a value obtained by dividing the added value by 2 to 100 is used as a keyword similarity and balanced with a cited document similarity described later. The keyword similarity calculation unit 92 stores the keyword similarity calculated in this way in the keyword similarity DB 85.

The stable inter-document distance calculation unit 22 includes a cited document evaluation value calculation unit 93 and a cited document similarity calculation unit 94. The cited document evaluation value calculation unit 93 has a function of referring to the cited document DB 86 and calculating an appearance rate of all cited documents cited as related documents in the key document in all reference documents. Specifically, the number of cited documents in all documents is calculated by dividing the number of documents that cite a cited document by the number of all reference documents. When the number of reference documents is N and the cited references are cited in C of all the documents, the intra-document cited document appearance rate is calculated as C / N. Furthermore, the cited document evaluation value calculation unit 93 has a function of calculating a cited document evaluation value by taking the reciprocal of the cited document appearance rate in all documents. That is, the cited document evaluation value is calculated by N / C, and indicates the degree (non-generality) in which each cited document is inherently included in the key document. It is conceivable to adjust the number of cited document appearances C to the power of 0.1 to 0.8. In this case, for example, when the adjustment to the power of 0.5 is performed, the cited document evaluation value is N / (C ^0.5 ).

  The cited document similarity calculation unit 94 adds the evaluation values for all documents cited in common by the key document and the search document, thereby calculating the cited document similarity indicating the degree of similarity between the key document and the search document. Has a function to calculate. The cited document similarity calculation unit 94 stores the cited document similarity calculated in this way in the cited document similarity DB 88.

  FIG. 19 is an example of the cited document DB 86. In the example of FIG. 19, the cited document similarity calculation unit 94 detects a document cited by duplication in the key document (US6713520B2) and the search document (US6433090B1). Three documents are cited in the key document (however, the document itself is also treated as a cited document). 23 documents are cited in the search document (however, the document itself is also handled as a cited document). Of these, only US6433090 (search document) is detected as a duplicate citation. The cited document evaluation value of this duplicate cited document becomes the cited document similarity between the key document and the search document. If the number of classification target documents exceeds 100, adjustment such as fixing the evaluation value to 1 is also conceivable.

  The stable inter-document distance calculation unit 22 includes a document similarity calculation unit 95 and an inter-document distance calculation unit 96. The document similarity calculation unit 95 has a function of calculating the similarity (document similarity) between the key document and the search document by combining the keyword similarity and the cited document similarity. As a specific example, the document similarity calculation unit 95 stores the average value of the keyword similarity and the cited document similarity as the document similarity in the document similarity DB 89. At this time, by dividing the keyword similarity by an appropriate number as described above, the balance between the strength of the influence of the keyword similarity on the stable inter-document distance and the strength of the influence of the cited document similarity on the stable inter-document distance. It is conceivable to adjust. Further, it is considered that each of the keyword similarity and the cited document similarity is divided by the sum of the number of keywords included in the key document and the search document and the number of cited references cited in the key document and the search document. It is done. In order to achieve this, the document similarity calculation unit 95 calculates the average value of the keyword similarity and the cited document similarity, and then calculates the average number of keywords included in the key document and the search document, the key document, It is conceivable that the document similarity is a value obtained by dividing the sum by the number of cited documents cited in the search document.

The inter-document distance calculation unit 96 has a function of calculating a stable inter-document distance L ₀ (i, j) between the document T _i and the document T _j using the similarity stored in the document similarity DB 89. . Here, the stable inter-document distance L ₀ (i, j) is expressed by the following equation:
L ₀ (i, j) = 2 / (S _ij + S _ji )
It is requested from. Here, S _ij represents the document similarity when the document T _i is the key document and the document T _j is the search document. S _ji is the document T _j is the key document and the document T _i is the search document. Represents the degree of document similarity. That is, the above expression means that the average value of the similarity calculated by switching the relationship between the key document and the search document is taken between the document T _i and the document T _j, and further, the inverse of the average value is taken. is doing. The _{reason why} the average value of similarity calculated by switching the relationship between the key document and the search document is used is that S _ij and S _ji do not necessarily match. The stable inter-document distance L ₀ (i, j) calculated in this way decreases as the similarity between the two documents increases, and increases as the similarity decreases.

  FIG. 8 is a flowchart showing a process for calculating a stable inter-document distance (a subroutine of the stable inter-document distance calculating step (S61) in FIG. 6). First, the keyword extraction unit 71 extracts keywords from the key document, and stores the extracted keywords in the keyword DB 81 (S801). Further, the reference word extraction unit 72 extracts the reference word from the reference document, and stores the extracted reference word in the all word DB 82 (S802). Next, the keyword evaluation value calculation unit 91 reads the reference word stored in the all word DB 82, calculates the evaluation value of each keyword, and stores the evaluation value in the keyword evaluation value DB 83 (S803). Further, the inter-document distance calculation unit 93 extracts a search word from the search document, and stores the extracted search word in the search word DB 84 (S804). Next, the similarity calculation unit 92 reads the evaluation value stored in the keyword evaluation value DB 83 and the search word stored in the search word DB 84, calculates the similarity between the key document and the search document, The similarity is stored in the keyword similarity DB 85 (S805).

  The cited document evaluation value calculation unit 93 extracts documents cited in common by the key document and the search document (S806), and calculates the cited document evaluation value of each of these documents (S807). The cited document similarity calculation unit 94 calculates the cited document similarity between the key document and the search document (S808). The document similarity calculation unit 95 calculates the document similarity by averaging the keyword similarity and the cited document similarity regarding the key document and the search document (S809). Finally, the inter-document distance calculation unit 93 calculates a stable inter-document distance between the key document and the search document based on the calculated document similarity (S810). When calculating the average of the keyword similarity and the cited document similarity, for example, by multiplying the cited document similarity by a coefficient of 2 to 40, preferably 8 to 11, weighting the cited document similarity, It is possible to focus on the evaluation based on the similarity of cited references between documents with similar contents, which causes fluctuations, and to evaluate based on the similarity of keywords between documents in different fields where there are no common cited references. By calculating the position coordinates based on the inter-document distance, it is possible to classify and arrange patent documents covering a wide technical field in more detail.

  Another example of stable document distance calculation processing will be described. In this example, a processing function (keyword evaluation value calculation unit 91 and keyword similarity calculation unit 92) for calculating a similarity based on a common keyword in the stable inter-document distance calculation unit 22 and a process for calculating a similarity based on a common cited document. Functions (cited document evaluation value calculation unit 93 and cited document similarity calculation unit 94) are integrated. In addition, the reference document table of this example is a combination of the data of all word DB 82 and the data of cited reference DB 86.

The calculation processing of this example will be described with reference to the flowchart of FIG. The stable document distance calculating unit 22 (keyword / cited document extracting unit 70) extracts words / cited documents from the key document and the search document (S211), and extracts and stores the common keywords and common cited documents (S212). . FIG. 21 shows data of common keywords and common citations extracted and stored. The stable inter-document distance calculation unit 22 refers to the reference document table, calculates the appearance rate of each common keyword and common cited document in the reference document table, and calculates / stores the evaluation value based on this (S213). The stable inter-document distance calculation unit 22 multiplies the evaluation value of each common cited document by a coefficient of 2 to 40, preferably 8 to 11, and replaces this value (S214). The stable inter-document distance calculation unit 22 calculates the document similarity by adding the keyword evaluation value obtained in S213 and the cited document evaluation value obtained in S214 (S215). Furthermore, it is preferable that the value obtained by dividing the total value by the total number of words / cited documents extracted in S211 is the document similarity. A stable inter-document distance is calculated based on the document similarity (S216).
In the above example, the similarity based on each is calculated by adding the evaluation values of the common keyword and the common cited document, but instead, the vector space method is applied to the common keyword similarity and the common cited document similarity. A method of calculating the document similarity after obtaining each and multiplying them by weight, and pre-weighting the evaluation values of the common keyword and the common cited document, and then obtaining the document similarity by the vector space method in a lump. Is also possible.
When the common keyword similarity is calculated by applying the vector space method, the processing shown in FIGS. 22 to 24 is performed as an example. Specifically, each element of the vector which is set to 1 when a word appears in the corresponding document shown in the columns (c) and (d) and 0 otherwise, and the evaluation value vector in the column (b) By performing multiplication, vectors in the columns (e) and (g) are calculated. As shown in columns (f) and (h), an operation for squaring each element of the vectors in columns (e) and (g) is performed, and further multiplied by a normalization coefficient to obtain a unit vector ( The unit vector (j) of the vectors in columns i) and (h) is calculated. The common keyword similarity can be calculated based on the distance or inner product between the unit vector (i) and the unit vector (j). The common cited document similarity can also be calculated by the same processing as described above. In addition to the keyword and the search word, the documents cited in the key document and the search document are simultaneously subjected to the vector space method calculation (in this case, the same input as in FIG. 21 is made in the column (a)). If the evaluation value in the column (b) is a value multiplied by a weight for balancing the common keyword similarity and the common cited document similarity, the document similarity is directly obtained by performing the above calculation thereafter. can do.

  FIG. 9 is a flowchart showing a subroutine of the total inter-document force vector calculation step (S66) of FIG. First, the inter-document force vector calculation unit 26 reads the position coordinates of each document stored in the position coordinate DB 16, and calculates the distance on the coordinate plane between the two documents (the length of the separation vector) from the position coordinates. (S91). The inter-document force vector calculation unit 26 reads the stable inter-document distance stored in the inter-document distance DB 14 and calculates the inter-document force using the stable inter-document distance and the distance calculated in the previous step S91. (S92). Further, the inter-document force vector calculation unit 26 calculates the X component and the Y component of the inter-document force based on the separation vector (S93), and calculates the sum of the inter-document forces working from other arranged documents for a certain document. By calculating as a vector sum, a total inter-document force vector is calculated (S94). Then, when the total document inter-force vector is calculated for all the arranged documents, the flow ends. When there is a document for which the total inter-document force vector is not calculated, the above steps (S91 to S94) are performed. Repeated (S95).

  FIG. 10 is a flowchart showing a subroutine of the position coordinate update step (S67) of FIG. First, the position coordinate update unit 28 reads an inter-document force vector stored in the inter-document force vector DB 18, and moves each document, that is, changes the position coordinate according to the vector (S101). Thereafter, the position coordinate update unit 28 calculates an average value of the moving distances of the respective documents used for determining the convergence condition (S102). In S <b> 68, the convergence condition can be that the average value of the movement distance is less than the threshold value. Alternatively, the convergence condition may be that the position coordinate update step is repeated int√N times.

  FIG. 11 is a flowchart showing processing for adding an additional arrangement document to the display coordinate system. Referring to FIG. 11, after the arrangement / movement process of the initial arrangement document shown in FIG. 6 is completed, additional arrangement documents are sequentially added to complete the arrangement / movement of all the classification target documents. Will be explained.

  The arrangement document selection unit 23 randomly selects the next arrangement document to be added to the display coordinate system int (mm / 10) (where mm is the number of classification target documents already arranged in the display coordinate system). (S111). However, when the remaining number of classification target documents is less than int (mm / 10), all the remaining classification target documents become time-arranged documents. The number of time-arranged documents added immediately after the arrangement / movement process of the initial arrangement document is int {(int√N) / 10}.

The position coordinate initial value setting unit 24 calculates the position coordinates on the display coordinate system where the time placement document is first installed (S112). In the present embodiment, the time-arranged document is initially set so as to be positioned in the vicinity of the arranged classification target document having the shortest stable inter-document distance. Specifically, the position coordinate initial value setting unit 24 refers to the stable inter-document distance DB 14 to obtain c where L ₀ (c, mm + k) is the minimum value (where c = 1 to mm). The initial value of the position coordinates of the arrangement document T _k is (X _k , Y _k ) = (X _c + ε, Y _c + ε) (ε: constant). Instead of the above, the initial value of the time-arranged document may be determined by a random number as in the case of the initial-arranged document.

  After the position coordinates of a certain time-arranged document are set to initial values, the inter-document force vector calculation unit 26 calculates the current total inter-document force vector received by the time-arranged document from other arranged classification target documents ( S113). The position coordinate update unit 28 updates the position coordinates of the time-arranged document based on the total inter-document force vector (S114). The convergence condition determination unit 30 determines the convergence of the position coordinates of the time-arranged document when the current movement amount of the time-arranged document is equal to or smaller than the threshold value or when the processes of S113 to S115 are executed a predetermined number of times ( S115). If convergence is not determined, the processing of S113 to S115 is repeated again for the time-arranged document.

  The processes of S112 to S115 are sequentially performed for other documents among the time-arranged documents. When the arrangement / movement processing is completed for all the documents in the time-arranged document, the position calculation (processing in S66 to S68) is performed for all the arranged documents including the time-arranged document added to the display coordinate system this time. This is performed N times (S117). In this way, all the time-arranged documents are arranged at once and moved (the process of calculating the total document inter-force vector and updating the position coordinates based on this is performed for each of all or some of the arranged documents. Instead of repeating the process until the convergence of the position is determined), the position coordinates of all the time-arranged documents are processed one by one by sequentially placing and moving the time-arranged documents one by one while fixing the position coordinates of the other arranged documents. , And the placement / movement process for all the arranged documents including the current time-arranged document is performed, so that the number of repetitions of the movement process can be reduced. The processing of S111 to S117 is performed on the unallocated classification target document. However, when there is no unallocated classification target document, the arrangement / movement process ends at this point.

  FIG. 12 is a flowchart showing a subroutine of the result display step (output step) in FIG. First, the display unit 32 divides the display area into m × n cells (S121). Here, the maximum value and the minimum value of the X coordinate and the Y coordinate defining the display area are input by the input unit 34 (S122). This input is performed by the user. Next, the display unit 32 calculates a coordinate range corresponding to each cell based on the above-described value input from the input unit 34 (S123). Then, the display unit 32 displays the number of documents having position coordinates within the coordinate range of each cell in the display area (S124). The display unit 32 creates an image of the document included in each cell (S125) and hyperlinks the image of the document to each cell (S126).

  FIGS. 13A and 13B are diagrams for explaining a modification of the result display by the display unit 32 of FIG. The display area 50 shown in the figure corresponds to FIG. In this example, by designating a part of the display area 50 as a new display area, the part can be displayed again on the entire display area 50. For example, if four cells in the center (the outer frame is indicated by a thick line) are designated in FIG. 13A, the designated portion is re-displayed in the entire display area 50 as shown in FIG. 13B. Is displayed. At this time, since the number of cells in the display area 50 is unchanged, the designated portion is divided into finer cells. For example, the cell in which the number of documents is displayed as “5” in FIG. 13A corresponds to the upper right four cells in FIG. Therefore, the sum of the number of documents in the four cells is 5. The designation of the display area can be performed, for example, by inputting coordinate values to the input unit 34 in FIG. 1 or by selecting with a mouse on the screen.

  FIG. 14 is a diagram for explaining a modification of the result display by the display unit 32 of FIG. In this example, the display unit 32 plots and displays the position coordinates of each document in the display coordinate system (two-dimensional display coordinate system) in the plot area 54p in the display area 54d. Each plot 56 has a text box 53 (“text box” having a certain shape and size for displaying the attributes of the plotted document so that the corresponding document can be identified. A small area, but it is also possible to set text boxes of different shapes or sizes depending on the display amount of attribute information.) In order to prevent the text box 53 from protruding from the display area 54d, the plot area 54p corresponds to the size of the text box 53 from the display area 54a (when a text box having a plurality of shapes or sizes is set, It has been reduced inward by the length of the long vertical or horizontal diameter. That is, the area surrounded by the dotted line located at each corner of the display area 54 a and the frame of the display area 54 a has the same size and shape as the text box 53. Here, an open patent gazette is assumed as the classification target document, and the application number is displayed as the document number. Further, a hyperlink is pasted on the displayed text box 53 or the document number, and the image of the document can be accessed by clicking the document number on the screen. In addition to the document number, the text box 53 can display the name of the invention, the applicant, and a keyword cut out from the summary, which makes it easier to grasp the classification contents.

  FIG. 15 shows a flowchart according to a modification of FIG. First, the user inputs the maximum value and the minimum value of the X coordinate and the Y coordinate that define the display area 54a (S151). This input can be performed from the input unit 34 of FIG. The display unit 32 sets the plot area 54p (S152). The display unit 32 plots the position coordinates of each document in the plot area 54p, and displays the text box 53 in association with the plot 56 (S153). Further, the display unit 32 hyperlinks the document image to each plot (S154). In the display example of FIG. 14, a partial area of the display is specified, and this is enlarged and displayed as a new display area, or one point of the display area, for example, a text box at the center of the area is specified with the mouse pointer. It is possible to display an enlarged / reduced display around this. Also, by copying the contents of one or more designated text boxes onto a worksheet such as spreadsheet software, the work following the classification can be made easier.

  FIG. 16 is a diagram for explaining a modification of the result display by the display unit 32 of FIG. In this example, the display unit 32 displays a document having a position coordinate within a display radius on the coordinate plane as a corresponding document with the position coordinate of the reference document specified by the user as a base point. Designation of the reference document and the display radius can be performed from the input unit 34 in FIG. The reference document is selected from the classification target documents stored in the classification target document DB 12. In FIG. 16, the corresponding document list is displayed on the right side in the display screen 57. Hyperlinks are pasted on these displays. Here, there are four documents within the display radius from the reference document. In addition, this list is sorted and displayed in order from the shortest distance from the reference document. Further, in this example, a display area 58 is provided on the left side in the display screen 57. In the display area 58, the position coordinates of each document are plotted with the reference document as the center, and a circle 59 having the display radius as the center with the reference document as a center is displayed. This circle 59 can be used as a guide when redesignating the display radius. The numbers attached to each plot correspond to the numbers in the corresponding document list. According to this example, a document similar to the reference document can be searched. In this case, the reference document is displayed in a so-called highlight display in which the color or font is changed, so that the positional relationship between the target reference document and a similar document can be easily understood.

  FIG. 17 shows a flowchart according to a modification of FIG. First, the user inputs a reference document and display deformation from the input unit 34 in FIG. 1 (S171, S172). Then, the display unit 32 reads the position coordinates of each document stored in the position coordinate DB 16, and displays the documents located within the display radius from the reference document as a corresponding document list (S173). Further, the display unit 32 hyperlinks the image of the document displayed in the corresponding document list (S174). Here, the user accesses the image of the document displayed by following the hyperlink as necessary, and confirms the content (S175). When the search radius is specified again and the search is performed again, the above steps (S172 to S175) are repeated, and when the search is not performed again, the flow is terminated (S176).

  Finally, the hardware configuration of the document classification device 1 will be described with reference to FIG. FIG. 18 is a block diagram showing a hardware configuration of the document classification device 1 of FIG. As shown in FIG. 18, the document classification device 1 physically includes a control device 1a, a memory 1b, a storage device 1c, an input device 1d, and a display device 1e. These devices are electrically connected to each other via a bus 1f so that various signals can be input and output.

  Specifically, the control device 1a is, for example, a CPU (Central Processing Unit), and the memory 1b is a volatile semiconductor memory such as a RAM (Random Access Memory). The storage device 1c is a non-volatile magnetic disk such as an HDD (Hard Disc Drive). The input device 1d is, for example, a keyboard or a mouse, and the display device 1e is an LCD (Liquid Crystal Display) or a CRT (Cathode Ray Tube) display.

  The correspondence between the hardware configuration and the functional configuration is shown below. Regarding the document classification device 1, the functions of the database 10 are realized by the storage device 1c as a physical component. Each function of the stable inter-document distance calculation unit 22, the position coordinate initial value setting unit 24, the inter-document force vector calculation unit 26, the position coordinate update unit 28, and the convergence condition determination unit 30 is executed by the control device 1a. It is realized by doing. Each function of the input unit 34 is realized by the input device 1d. In addition, each function which the display part 32 has is implement | achieved by the control apparatus 1a and the display apparatus 1e. That is, the control device 1a performs a predetermined calculation to determine the display content of the classification result, and the display device 1e displays the classification result according to the content.

  The document classification apparatus and the document classification method according to the present invention are not limited to the above-described embodiments, and various modifications can be made. For example, although the configuration in which the position coordinates of each document are determined on a two-dimensional coordinate plane is shown, the coordinate plane may be one-dimensional. At this time, each document has position coordinates on one straight line. In this case, too, it is referred to as a one-dimensional “coordinate plane” for convenience. Alternatively, the position coordinates of each document may be determined by extending to three or more dimensions.

  Further, although the convergence condition is that the average value of the moving distance of each document is equal to or less than a specified value, the convergence condition is not limited to this. For example, the convergence condition may be that the maximum value of the moving distance of each document is not more than a specified value.

  Further, the movement coefficient k used for updating the position coordinates does not always need to be a constant value. After the convergence has progressed to some extent, the movement coefficient k may be adjusted depending on whether the average value of the movement distance of each document is increased or decreased in order to increase the convergence speed. For example, if the average value of the movement distance is larger than that after the previous update, k ′ = k × 0.01 (k ′: movement coefficient after addition / subtraction), and if it is smaller, k ′ = k × 1.03.

  It can show so that the relationship of the similarity between several patent documents can be recognized visually.

It is a block diagram which shows one Embodiment of the document classification device by this invention. It is a block diagram which shows an example of the database of inter-document distance DB14. It is a block diagram which shows an example of the database of position coordinate DB16. It is a block diagram which shows an example of the database of document force vector DB18. It is a figure which shows an example of the result display screen by the display part. It is a flowchart which shows the process which arrange | positions and moves an initial arrangement document in an initial process and a two-dimensional display coordinate system. FIG. 2 is a block diagram illustrating an example of a configuration of a stable inter-document distance calculation unit 22 in FIG. 1. It is a flowchart which shows operation | movement of the stable document distance calculation part 22 of FIG. It is a flowchart which shows the subroutine of the total document inter-force vector calculation step (S66) of FIG. It is a flowchart which shows the subroutine of the update step (S67) of the position coordinate of FIG. It is a flowchart which shows the process which adds an additional arrangement | positioning document to a display coordinate system. It is a flowchart which shows the subroutine of the result display step of FIG. (A) And (b) is a figure for demonstrating the modification of the result display by the display part 32 of FIG. It is a figure for demonstrating the modification of the result display by the display part of FIG. 15 shows a flowchart according to a modification of FIG. It is a figure for demonstrating the modification of the result display by the display part of FIG. FIG. 17 shows a flowchart according to a modification of FIG. It is a block diagram which shows the hardware constitutions of the document classification device 1 of FIG. It is an example of cited reference DB86. The flowchart in the 2nd example of the calculation process of the distance between stable documents is shown. The data of the common keyword extracted in the 2nd example of the calculation process of the distance between stable documents, and the common cited reference data are shown. It is a 1st figure explaining embodiment which applied the vector space method. It is a 2nd figure explaining embodiment which applied the vector space method. It is a 3rd figure explaining embodiment which applied the vector space method.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Document classification apparatus, 10 ... Database, 12 ... Classification object DB, 14 ... Inter-document distance DB, 16 ... Position coordinate DB, 18 ... Inter-document force vector DB, 22 ... Stable inter-document distance calculation part, 23 ... Arrangement Document selection unit, 24 ... Position coordinate initial value setting unit, 26 ... Inter-document force vector calculation unit, 28 ... Position coordinate update unit, 30 ... Convergence condition determination unit, 32 ... Display unit, 34 ... Input unit, 53 ... Text box 54a ... display area, 54p ... plot area, 56 ... plot, 91 ... keyword evaluation value calculation unit, 92 ... keyword similarity calculation unit, 93 ... cited document evaluation value calculation unit, 94 ... Cited reference similarity calculation unit, 95... Document similarity calculation unit, 96.

Claims (13)

A document classification method for classifying a plurality of documents to be classified according to their contents,
A common keyword extraction step of extracting a common keyword that appears in common among two of the classification target documents;
A common cited reference extracting step of extracting a common cited reference that is commonly cited in two of the classification target documents;
The number of occurrences of the common keyword in the keyword table is counted by referring to the keyword table in which the word included in each keyword reference document is stored in association with the identification number of the keyword reference document. A common keyword evaluation value calculating step for calculating a keyword non-generality evaluation value;
Referring to the cited document table that stores the documents cited in each cited document reference document in association with the identification number of the cited document reference document, the number of appearances of the common cited document in the cited document table is counted, and this appearance The common cited reference evaluation value calculating step of calculating an evaluation value of non-generality of the common cited reference based on a number;
A keyword similarity calculating step of adding the evaluation values of the common keywords and calculating a similarity based on the common keywords of two documents of the classification target documents;
Cited document similarity calculation step of calculating the similarity based on the common citations of two documents of the classification target documents by adding the evaluation values of the common citations;
A document similarity calculation step for calculating a document similarity by combining the similarity based on the common keyword and the similarity based on the common cited document;
A document moving step in which the classification target documents are arranged in a coordinate system, and each of the classification target documents is moved so that the two documents with high document similarity approach and the two documents with low document similarity move away from each other. A document classification method characterized by comprising: 2. The document classification method according to claim 1, wherein in the document similarity calculation step, an average value of the similarity based on the common keyword and the similarity based on the common cited document is set as the document similarity. 2. The document similarity calculation step, wherein an average value of a value obtained by dividing a similarity based on a common keyword by a numerical value of 2 to 100 and a similarity based on a common cited document is used as the document similarity. Document classification method described in 1. 4. The document classification method according to claim 3, wherein, in the common keyword evaluation value calculating step, a value obtained by multiplying the number of appearances of the common keyword by a power of n (n = 0.1 to 0.8) is substituted. In the common keyword evaluation value calculating step, the number of occurrences of the common keyword is replaced with a value obtained by raising n (n = 0.1 to 0.8).
In the common cited document evaluation value calculation step, the number of appearances of the common cited document is replaced with a value obtained by raising n (n = 0.1 to 0.8).
In the document similarity calculation step, a sum of a value obtained by dividing a similarity based on a common keyword by a numerical value of 2 to 100 and a similarity based on a common cited document is obtained, and the value is calculated as two documents of the classification target documents. 2. The document classification according to claim 1, wherein a value obtained by dividing the number of words included in a document by the sum of the number of documents cited in two of the classification target documents is used as the document similarity. Method. The document classification method according to claim 1, wherein the document is regarded as a document cited in the document. The document classification method according to claim 1, wherein the classification target document is a patent document. 2. The document classification method according to claim 1, wherein, when counting the number of appearances of the common keyword, the same word is not counted repeatedly as being associated with an identification number of a certain keyword reference document. A document classification method for classifying a plurality of documents to be classified according to their contents,
A common keyword similarity calculating step of calculating a similarity between the two documents based on a common keyword that appears in common among the two documents of the classification target document;
A common cited reference similarity calculating step for calculating a similarity between the two documents based on a common cited reference cited in two documents of the classification target documents;
A document similarity calculation step for calculating a document similarity by combining the similarity based on the common keyword and the similarity based on the common cited document;
A document moving step in which the classification target documents are arranged in a coordinate system, and each of the classification target documents is moved so that the two documents with high document similarity approach and the two documents with low document similarity move away from each other. A document classification method characterized by comprising: The document classification method according to claim 9, wherein a vector space method is applied in the common keyword similarity calculation step, the common cited document similarity calculation step, and the document similarity calculation step. A document classification program for causing a computer system to execute the steps according to any one of claims 1 to 10. A document classification device for classifying a plurality of classification target documents according to their contents,
Common keyword extraction means for extracting common keywords that appear in common in two of the documents to be classified;
Common citation extraction means for extracting common citations that are commonly cited in two of the classification target documents;
The number of occurrences of the common keyword in the keyword table is counted by referring to the keyword table in which the word included in each keyword reference document is stored in association with the identification number of the keyword reference document. A common keyword evaluation value calculating means for calculating a keyword non-generality evaluation value;
Referring to the cited document table that stores the documents cited in each cited document reference document in association with the identification number of the cited document reference document, the number of appearances of the common cited document in the cited document table is counted, and this appearance The common cited reference evaluation value calculating means for calculating an evaluation value of non-generality of the common cited reference based on a number;
A keyword similarity calculating means for adding the evaluation values of the common keywords and calculating a similarity based on the common keywords of two documents of the classification target documents;
Cited document similarity calculation means for calculating the similarity based on the common citations of two documents of the classification target documents by adding the evaluation values of the common citations;
A document similarity calculating means for calculating a document similarity that combines the similarity based on the common keyword and the similarity based on the common cited document;
Document moving means for performing processing for moving the classification target documents so that the two documents with high document similarity approach each other and the two documents with low document similarity move away from each other, arranged in the coordinate system. A document classification device comprising: A document classification device for classifying a plurality of classification target documents according to their contents,
Common keyword similarity calculating means for calculating the similarity between the two documents based on a common keyword that appears in common among the two documents to be classified;
A common cited reference similarity calculating means for calculating the similarity of the two documents based on a common cited reference that is commonly cited in two of the classification target documents;
A document similarity calculating means for calculating a document similarity that combines the similarity based on the common keyword and the similarity based on the common cited document;
Document moving means for performing processing for moving the classification target documents so that the two documents with high document similarity approach each other and the two documents with low document similarity move away from each other, arranged in the coordinate system. A document classification device comprising:

Images