JP2010020530A - Document classification providing device, document classification providing method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a document classification providing device, a document classification method and a program, allowing reduction of providing leakage of classification to a document, and allowing providing of high-reliability classification. <P>SOLUTION: A characteristic amount extraction part 12 extracts the IPC (International Patent Classification) from a patent publication input from a document input part 11. A document characteristic amount vector generation part 13 extracts the head IPC from the IPCs obtained in the characteristic amount extraction part 12 in reference to a main characteristic amount dictionary 14, weights the head IPC, and thereafter generates a document vector with the IPCs as elements. An inter-document distance calculation part 15 calculates a distance between the patent publications by use of the document vectors. A two-dimensional coordinate mapping processing part 16 arranges a coordinate point corresponding to each patent publication on two-dimensional coordinates such that relation of the distance between the patent publications appears. A classification non-providing subset selection part 17 sets a circular area including a classification providing patent publication having the head IPC on the two-dimensional coordinates, and selects a set of the coordinate points corresponding to the classification non-providing patent publications inside the circular area. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a document classification assigning device, a document classification assigning method, and a program for assigning a predetermined classification to an input document, and the classification classification is based on a document to which a predetermined classification classification is assigned. The present invention relates to a document classification assignment apparatus, a document classification assignment method, and a program for assigning a classification category to a document that has not been assigned and presenting a search formula for obtaining a document to which the predetermined classification category is to be assigned. The present invention relates to a document classification assigning apparatus, a document classification assigning method, and a program for selecting a patent publication related to a predetermined technical classification using a dimension map, assigning a classification classification, and presenting a search formula for obtaining the patent publication .

  Conventionally, methods for automatically classifying input documents by some index have been proposed. However, complete automatic classification cannot be achieved by these methods. Even if automatic classification was performed, it was necessary for a person to check and judge the contents. As a technique for automatically classifying and searching documents, there is, for example, Patent Document 1. The technique described in Patent Literature 1 narrows down search targets by item constituting a sentence in document search, and by narrowing down the search targets, search according to the purpose is possible and search accuracy is improved. .

Further, the classification of documents is generally performed based on the contents described in the documents. Japanese Patent Application Laid-Open No. 2004-228561 describes a technique that allows the relationship between content similarities to be visually recognized by expressing the content similarity between patent gazettes as a distance. This technique is a technique for displaying a plurality of documents on two-dimensional coordinates in accordance with the contents of each of the documents. The inter-document distance between each two documents in the plurality of documents is obtained, and each of the documents is displayed on the two-dimensional coordinates. The plurality of documents are arranged so as to reproduce the magnitude relationship of the distance between the two documents. Documents having similar contents are arranged close to each other, and the whole of the plurality of documents can be visually recognized. Therefore, it is possible to give an overview of the entire document set. In the case of assigning classification categories, there is an advantage that classification assignment can be easily performed because there are similar contents nearby.
JP-A-11-296550 JP 2006-190235 A

  As an index for determining whether or not the contents of a document are similar, there are words used in the document, but even if different words are used, the same contents are often expressed. Even if the same word is used, it often represents different contents. In Patent Document 1, since it is mechanically determined whether or not the documents represent the same content depending on the presence or absence of the specified keyword included in the document, even if the content is similar, if there is no specified keyword, it is reflected in the search result. There was a problem that a search omission occurred. In addition, even if a document including the specified keyword is obtained, there is a problem that search noise that the contents are different is mixed. Further, the content of the document cannot be determined only by the presence or absence of a word, and a step of re-evaluating the obtained result is necessary. However, the evaluation method for the obtained result is referred to in Patent Document 1. No.

  Further, in Patent Document 2, since a document set is arranged on two-dimensional coordinates based on the similarity between documents, there is a problem that the boundary between what is similar and what is not similar is not clear. is there. Document A and document B are located close to each other on the two-dimensional coordinate, document B and document C are located close to each other on the two-dimensional coordinate, and document A and document B, document B and document C are Even when similar contents are held, the contents of document A and document C are often not similar. FIG. 13 shows an example of character discrimination when such a problem occurs. In the drawing, the character 111 represents Katakana “ya”, and the character 113 represents Katakana “ka”. Character 112 represents any one or another symbol. The character 111 and the character 112 are both highly similar as representing the contents of “Y”, and are arranged close to each other on the two-dimensional coordinates. The character 112 and the character 113 are both highly similar as representing the contents of “K”, and are arranged close to each other on the two-dimensional coordinates. However, the characters 111 and 113 are “YA” and “F”, and the contents are completely different. Although katakana notation is shown in FIG. 13 as an example, the same problem arises with respect to the similarity of document contents. For example, when document A, document B, and document C exist, it is unclear whether a boundary line should be drawn between A-B or B-C from the viewpoint of content similarity. Furthermore, there is a problem that such a relationship between the document A, the document B, and the document C cannot be determined from a diagram arranged on two-dimensional coordinates.

  Therefore, the present invention has been made paying attention to the above-mentioned conventional unsolved problems, and it is possible to reduce the omission of classification classification to a document and to provide a classification classification with high reliability. An object is to provide a granting device, a document classification method, and a program.

  In order to solve the above problem, the document classification assigning apparatus according to claim 1 of the present invention is provided with classification categories based on a classification grant document that is a document with predetermined classification categories. A document input unit that inputs a plurality of documents including the classified document and the non-classified document in the document classification and grant device that assigns the predetermined classification category to a non-classified document that is a non-classified document; and A feature amount extraction unit that extracts a feature amount from each document input from a document input unit; a main feature amount dictionary storage unit that stores a rule for extracting a main feature amount from the feature amount; and the main feature amount dictionary The feature after extracting a main feature amount from the feature amount obtained from the feature amount extraction unit with reference to a rule stored in the storage unit and superimposing a weight greater than 1 on the value of the main feature amount Special features of quantity A document feature vector generation unit that generates a quantity vector, an inter-document distance calculation unit that calculates a distance between documents using the feature vector generated by the document feature vector generation unit, and the inter-document distance calculation unit The two-dimensional coordinate mapping processing unit that arranges the coordinate points corresponding to each document on the two-dimensional coordinates, and the main feature amount from the two-dimensional coordinates so that the relationship between the distances between the documents obtained in step 1 can be expressed. And a non-classified subset selection unit that selects a set of coordinate points corresponding to the non-classified document to which the predetermined classification category should be assigned. To do.

  According to the first aspect of the present invention, the distance between the documents is calculated after weighting the main feature amount of each document, and the correspondence between the distances between the documents is represented on the two-dimensional coordinates. Since the coordinate points are arranged, it is possible to arrange the document to be classified and the other documents so that they are separated, and the classification category should be assigned to the unclassified documents that are close to the classified document By selecting as an unassigned document, it is possible to assign a classification category with high reliability with few omissions.

  According to a second aspect of the present invention, there is provided the document classification assigning apparatus according to the first aspect, wherein the filtering feature amount is extracted from the classification imparted document input from the document input unit and used for determining the similarity of the document. For each of the predetermined classification categories, a feature amount extraction unit, a shape of a frequency distribution of filtering feature amounts included in the entire classification-added document, and a frequency distribution of filtering feature amounts included in each of the classification-added documents And the frequency distribution of the filtering feature quantity of the predetermined classification-added document based on the determined similarity degree is the frequency of the filtering feature quantity of the entire classification-added document. It is a case where it is determined that the shape of the distribution is not similar, and the highest frequency in the frequency distribution of the filtering feature quantity possessed by the entire classified document When the high filtering feature quantity and the filtering characteristic quantity with the highest frequency in the frequency distribution of the filtering feature quantity of the predetermined classification assignment document are not the same, the predetermined classification assignment document A classification removal unit for removing the classification assigned to the document, and in order to place coordinate points corresponding to the document obtained by the classification removal unit on the two-dimensional coordinates, Input to the mapping processing unit.

According to the second aspect of the present invention, the similarity with respect to the entire frequency distribution of the filtering feature quantity possessed by each classified document is determined, and it is determined that the similarity of the predetermined classified document is low and not similar. In this case, the assigned classification category is not valid and the reliability may be low. Therefore, by removing the classification category from the given classification-assigned document, the classification-assigned document set to which the high-reliability classification category is assigned Is obtained. Therefore, it is possible to select an appropriate document to which a classification category is to be assigned using the classification-assigned document set, and it is possible to assign a highly reliable classification category.
According to a third aspect of the present invention, there is provided the document classification assigning apparatus according to the second aspect, wherein the classification removing unit uses a chi-square test process to determine the similarity of the shape of the frequency distribution of the feature quantity for filtering. And

According to a fourth aspect of the present invention, there is provided the document classification assigning apparatus according to any one of the first to third aspects, wherein the unclassified subset selection unit is a coordinate point corresponding to the classified document on the two-dimensional coordinates. And a set of coordinate points corresponding to the unclassified document that is included in the circular region and has the main feature amount is selected.
According to the fourth aspect of the present invention, the classification-unassigned document to be classified is a non-classified document having the main feature amount, so that the coordinate point corresponding to the document to be classified and the other 2 It is possible to obtain a boundary on a dimensional coordinate, and by narrowing down the coordinate points corresponding to the classification-unassigned document to be classified to a circular area including the coordinate points corresponding to the classification-assigned document, it is efficient. Classification can be given.

According to a fifth aspect of the present invention, there is provided the document classification assigning apparatus according to any one of the first to fourth aspects, wherein the plurality of inputted documents are patent gazettes, and the feature amount and filtering for the extracted document are used. The feature quantity is a word in IPC (International Patent Classification), FI, FT (File Forming Term) or patent gazette, and the classification category is a technical classification category, and the main document feature The rule stored in the quantity dictionary storage unit is to extract a first IPC, an FI similar in content to the first IPC, an FT related to the FI, or a word in a patent gazette.
According to the fifth aspect of the present invention, it is possible to apply a technical classification with high reliability with little omission to a patent publication.

  In addition, the document classification method according to the invention of claim 6 is based on a classification-assigned document that is a document to which a predetermined classification category is assigned, and the previously classified document to which no classification category is assigned is assigned to the document classification method. In a document classification assigning method performed by a document classification assigning device for assigning a predetermined classification, a document input step for inputting a plurality of documents including the classified grant document and the unclassified document, and an input from the document input step A feature amount extracting step for extracting a feature amount from each document, a main feature amount dictionary storing step for storing a rule for extracting a main feature amount from the feature amounts in a main feature amount dictionary storage unit, and the main feature The feature amount after extracting a main feature amount from the feature amounts obtained in the feature amount extraction step with reference to the amount dictionary storage unit and superimposing a weight greater than 1 on the main feature amount A document feature vector generation step for creating a feature vector as an element, an inter-document distance calculation step for calculating a distance between documents using the feature vector generated in the document feature vector generation step, and the document A two-dimensional coordinate mapping step of arranging coordinate points corresponding to each document on the two-dimensional coordinates so that the relationship of the distance between documents obtained in the inter-distance calculation step is expressed; A non-classified subset selection step of selecting a coordinate point corresponding to a non-classified document that is a non-classified document having a main feature amount and to which the predetermined classification category should be assigned. Features.

  A document classification method according to a seventh aspect of the present invention is the document classification method according to the sixth aspect, wherein the filtering feature amount used for the similarity determination of the document is extracted from the classified document input from the document input step. Feature amount extraction step; for each of the predetermined classification categories, a filtering feature amount frequency distribution of the entire classified document, and a filtering feature amount frequency distribution of each of the classified documents The frequency distribution of the filtering feature quantity of the predetermined classification-added document based on the determined similarity degree is the shape of the frequency distribution of the filtering feature quantity of the entire classification-added document. In the frequency distribution of the filtering feature quantity of the entire classification-added document. If the feature quantity for filtering having a high degree is not the same as the feature quantity for filtering having the highest frequency in the frequency distribution of the feature quantity for filtering included in the predetermined classification-added document, the predetermined classification A classification removal step for removing a classification section assigned to the assigned document, and in the two-dimensional coordinate mapping processing step, a coordinate point corresponding to the document obtained in the classification removal step is placed on the two-dimensional coordinates. It is characterized by arranging.

The document classification method according to an eighth aspect of the present invention is the document classification method according to the seventh aspect, wherein, in the classification removal step, a chi-square test process is used for determining the similarity of the shape of the frequency distribution of the filtering feature quantity. Features.
A document classification method according to a ninth aspect of the present invention is the document classification method according to any one of the sixth to eighth aspects, wherein, in the unclassified subset selection step, coordinates corresponding to the classified document on the two-dimensional coordinates. A circular area including points is set, and a set of coordinate points corresponding to the unclassified document included in the circular area and having the main feature amount is selected.

  The document classification method according to a tenth aspect of the present invention is the document classification method according to any one of the sixth to ninth aspects, wherein the plurality of input documents are patent gazettes, and the feature amount of the extracted document and the filtering Is an IPC (International Patent Classification), FI, FT (File Forming Term) or a word in a patent gazette, the classification category is a technical classification category, and the main document The rule stored in the feature dictionary storage unit is to extract the first IPC, the FI similar in content to the first IPC, the FT related to the FI, or a word in the patent publication.

A program according to an eleventh aspect of the invention is a program for causing a computer to execute the method according to any one of the sixth to tenth aspects.
According to invention of Claim 11, this program is memorize | stored in a computer by the download from a server or the copy from a recording medium, and is made to perform, The method as described in any one of Claim 6 thru | or 10 is made into a computer Can be realized.

  According to the present invention, the main feature amount of the document is weighted, the distance between the documents is represented on the two-dimensional coordinates, and the unclassified document to which the classification category is to be assigned is selected. It is possible to assign classification categories with high reliability.

  Embodiments of the document classification assigning apparatus of the present invention will be described below with reference to the drawings.

  In this embodiment, among a set of patent gazettes obtained by searching for a specific technology, a set of classification grant patent gazettes to which a classification of technology classification (hereinafter simply referred to as “technical classification”) is assigned, and a technology A case will be described in which a set of unclassified patent gazettes to which no classification is assigned is input to the document classification assignment apparatus. Here, the technical classification assigned in advance is, for example, “O” or “X” (belongs to or does not belong), which is whether or not it belongs to a specific technical field.

  FIG. 1 is a configuration diagram of a document classification assigning apparatus according to the present embodiment. As shown in the figure, the document classification assignment apparatus includes a document input unit 11, a feature amount extraction unit 12, a document feature amount vector generation unit 13, a main feature amount dictionary (corresponding to a “main feature amount dictionary storage unit”) 14. The inter-document distance calculation unit 15, the two-dimensional coordinate mapping processing unit 16, and the unclassified subset selection unit 17 are provided. The main feature dictionary 14 is a database provided in a storage device (not shown) included in the document classification assigning device, and includes a document input unit 11, a feature extraction unit 12, a document feature vector generation unit 13, and an inter-document distance calculation unit. 15, the two-dimensional coordinate mapping processing unit 16 and the unclassified subset selection unit 17 are functions realized by a CPU (not shown) included in the document classification assigning device executing a program stored in the storage device. .

The document input unit 11 receives an unclassified patent gazette set 100 and a classified grant patent gazette set 101.
The feature quantity extraction unit 12 extracts an IPC (International Patent Classification) as a feature quantity of each patent publication.
The document feature quantity vector generation unit 13 generates a document vector having each IPC extracted by the feature quantity extraction unit 12 as an element and having the appearance frequency as a value for each patent publication. At that time, the main feature dictionary 14 is referred to, and the value corresponding to the first IPC is multiplied by ten. Here, the first IPC is a classification representing the invention described in the patent gazette, and is generally described at the top of the IPC description column of the patent gazette.

  The main feature dictionary 14 stores rules for extracting main features from the main features (first IPC in this embodiment) and feature values extracted from the patent gazette input from the document input unit 11. Has been. In this embodiment, a rule for extracting the leading IPC from the IPC described in the patent publication is stored. Further, the main feature dictionary 14 stores a rule for weighting the extracted first IPC 10 times.

The rule for extracting the main feature amount is not limited to the first IPC, but an FI similar in content to the first IPC, an FT (File Forming Term) related to the FI, a word in the patent gazette, etc. May be a rule for extracting.
The inter-document distance calculation unit 15 calculates the distance between each two documents from the generated document vector of each document. In this embodiment, the cosine distance is used as the inter-document distance. As shown in FIG. 2, the cosine distance is obtained by 1-cos (θ) based on an angle (θ) formed by two document vectors. When the document vectors of two documents are the same, the distance is 0, and when the document vectors of these two documents are arranged on the two-dimensional coordinates, they are arranged at the same coordinates.

  The two-dimensional coordinate mapping processing unit 16 maps the coordinate points corresponding to each patent publication on the two-dimensional coordinates so as to reproduce the magnitude relationship of the cosine distance, and the classification grant patent publication to which the technical classification is assigned in advance. The coordinate points corresponding to the patent publications included in the set 101 are highlighted in the two-dimensional coordinates. The value of the coordinate point corresponding to each patent publication is obtained by a known multidimensional scale construction method. FIG. 3 is a patent publication two-dimensional coordinate mapping diagram in which coordinate points corresponding to the obtained patent publications are displayed on the two-dimensional coordinates. In the figure, each point corresponds to one patent gazette, and a black circle corresponds to a highlighted patent gazette in which the affiliation to a specific technical classification is ○ (genus). By using a feature vector obtained by multiplying the weight of the first IPC by 10, the figure in which patent publications having the same first IPC are close to each other is obtained. FIG. 4 is a patent gazette two-dimensional coordinate mapping diagram in the case where the leading IPC is not weighted when the document vector is created. In the figure, each point corresponds to one patent gazette, and a black circle corresponds to a patent gazette in which a particular technical classification belongs to ○ (genus). In FIG. 4, only the relationship between the documents is expressed, and it is clear that the semantic boundary is unknown. Then, by comparing FIG. 3 and FIG. 4, it is clear that the semantic boundary between the documents becomes clear when the first IPC is weighted. Note that the weighting of the first IPC is not limited to 10 times, and the same effect can be obtained by weighting with a value larger than 1.

  The unclassified subset selection unit 17 selectively extracts coordinate points corresponding to unclassified patent gazettes to which a specific patent classification should be assigned centering on a portion surrounded by a black circle. At that time, the unclassified subset selection unit 17 refers to the main feature dictionary 14 and sets a coordinate point to be extracted to a patent publication having the same first IPC as the first IPC stored in the main feature dictionary 14. By limiting to the corresponding coordinate points, selection is performed in consideration of the boundary of the selection range. FIG. 5 shows a part of a circular area selected as an area for extracting coordinate points corresponding to an unclassified patent publication to which a specific patent classification should be assigned. For patent gazettes corresponding to coordinate points arranged in the circled area “A”, it is highly probable that the technical classification will be given, and the technical classification is given by selecting it as a classification grant target. Leakage can be effectively reduced. Furthermore, using the classification grant patent gazette and the classification non-grant patent publication included in the region for extracting the coordinate points corresponding to the non-classification patent gazette, the IPC that appears frequently in the patent gazette is extracted, and the high frequency By adding the appearing IPC to the search condition, it is possible to obtain a search expression that can obtain a search result with higher accuracy.

  FIG. 7 is a diagram showing the accuracy of classification assignment and non-assignment in the areas “A” and “A” shown in FIG. The accuracy (accuracy rate) shown in the figure is the accuracy obtained based on the number of correct patent gazettes in which the affiliation to a specific technology in each region is ○ (genus). In the area “A”, the correct patent gazette is extracted with an accuracy of 35%, and the accuracy of the unassigned rate of 100% is obtained in the area “a”, which is an area not to be selected. A low classification imparting accuracy in the region “A” indicates that there is a lot of search noise, but on the other hand, it suggests that there are few search omissions. Further, the accuracy of unassigned classification in the area “A” is 100% means that there is no correct patent gazette in the area, and the area “A” may be excluded from the extraction target at the time of assigning the classification. This shows the effectiveness of this method.

  In addition, effective publication selection can be performed by using a ratio in which the genus is ◯ (genus) in the classification grant patent publication set 101 instead of the frequency. FIG. 8 shows a three-dimensional display in which the number of patent publications is added to the patent publication two-dimensional coordinate mapping diagram as the z-axis. In the figure, dark regions are regions where the ratio of patent gazettes where the genus is ◯ (genus) is high. The three-dimensional area “A” in FIG. 8 corresponds to the two-dimensional area “A” in FIG. The region “α” represents a region other than the region “A” among regions where coordinate points corresponding to patent gazettes having the same first IPC exist. The coordinate points corresponding to the patent publication having the same first IPC are concentrated in the region “α”, but the coordinate points corresponding to the patent publication having the genus of ○ (genus) are concentrated in the region “A”. By selecting only the region “A” by the unclassified subset selection unit 17, it is possible to effectively select a patent publication to which a specific patent classification is assigned.

FIG. 9 is a flowchart showing the operation of the unclassified subset selection unit 17.
In S1301, the storage area of the unclassified document set to be output at the end of this process is initialized.
In S1302, it is determined whether or not processing has been performed for all patent gazettes with a genus of ○ (genus). The classification unassigned document set is output as the classification grant target patent gazette subset 18 to which the above patent classification is to be assigned. On the other hand, if all the attributes have not been processed for the patent gazette of ○ (genus), the process proceeds to S1303.

In S1303, one that has not been processed is selected from among patent gazettes with a genus of ○ (genus). In S1304, a circular area having a predetermined radius Rmax is set around the selected coordinate point on the two-dimensional coordinate of the above-mentioned patent gazette. In S1305, whether or not the circle belongs to the circle area (genus). It is determined whether there is any patent gazette. If there is no patent gazette of which genus is ○ (genus) inside the circular region, the coordinate value of the coordinate point on the two-dimensional coordinate of the patent gazette is set in the variable o meaning the center of the circle in S1306, A predetermined Rmin is set in a variable r indicating a radius. On the other hand, if there is a patent publication in which the genus is ◯ (genus) inside the circular region, each of the genus in the circular region is ◯ (genus) in the variable o meaning the center of the circle in S1307. The center of gravity value of the coordinate point corresponding to the patent gazette is set, and the variable r meaning the radius is set to the minimum value including all the coordinate points corresponding to the patent gazette of ○ (genus) in the circular region. To do.
In S1308, the unclassified patent publication included in the circle formed by the variable o indicating the center of the circle and the variable r indicating the radius is added to the unclassified subset.

In the second embodiment, a classification removal process is added to the first embodiment. FIG. 10 shows a configuration diagram of the document classification assigning apparatus according to the present embodiment. The document classification assigning apparatus according to the second embodiment is different from the document classification assigning apparatus according to the first embodiment in that a classification addition document input unit 20, a filtering feature amount extraction unit 21, and a classification removal unit 22 are added. This is different from the first embodiment. Other than that is the same as Example 1.
The filtering feature quantity extraction unit 21 extracts an IPC as a filtering feature quantity to be used for similarity determination from the classification grant patent publication input from the classification grant document input unit 20.

  The classification removing unit 22 extracts the IPC histogram of the entire patent publication set for the patent publication set whose genus is ◯ (genus), and then, in the patent publication set whose genus is ○ (genus). For each patent publication of No. 1, the similarity between the IPC histogram of the patent publication and the IPC histogram of the entire patent publication set is determined, and the similarity is not similar, for example, lower than the reference value And the patent gazette thus obtained by the classification removing unit 21 is input to the two-dimensional coordinate mapping processing unit 16, and the two-dimensional coordinate mapping processing unit 16 performs the classification removal. Coordinate points corresponding to the patent gazette input from the unit 21 are arranged on the two-dimensional coordinates.

FIG. 11 is a flowchart showing the operation of the classification removal unit 21.
In S71, a histogram (overall histogram) representing the appearance frequency of the IPC possessed by the entire patent gazette whose genus is ◯ (genus) as the investigation target is created.
In S72, it is determined whether or not all the patent publications to be searched have been searched. If the search has been completed, the process ends. If the search has not been completed, one unsearched patent publication is selected in S73.
In S74, the IPC histogram (individual histogram) of the patent publication is created.
In S75, the similarity of distribution between the whole histogram and the individual histogram is measured. In this embodiment, a chi-square value is used.

In S76, the chi-square value obtained in S75 is determined at a significance level of 5%. If it is determined that the distribution of the whole histogram and the individual histogram is similar, the process proceeds to the next patent publication. . On the other hand, if it is determined that they are not similar, in step S77, the IPC having the highest frequency in the entire histogram is compared with the IPC having the highest frequency in the individual histogram. And proceed to processing of the next patent candidate. On the other hand, when the IPCs having the highest frequency are different from each other, it is determined that the assigned technical classification is suspicious, and the technical classification assigned to the patent gazette is removed in S78, and the next patent publication is processed. Proceed with
By the above process, the technical classification that seems to be suspicious is removed, and a collection of patent gazettes to which a highly reliable technical classification is assigned is obtained, so that effective selection of patent publications to which a technical field should be assigned is realized.

The third embodiment is different from the first embodiment in that a process for dividing the size of the entire unclassified patent publication set 100 to which no technical classification is assigned is added to the input document. All other processing is equivalent. FIG. 12 is a configuration diagram of the document classification assigning apparatus according to the present embodiment.
The unclassified document set dividing unit 91 divides the inputted unclassified patent gazette set 100 into three, and each of the divided non-granted patent gazette sets 92 together with the classified grant patent gazette set 101 is a document input unit 11. To enter. This embodiment reduces the processing time required for the two-dimensional coordinate mapping process. A non-classified patent gazette to which a classification is assigned is obtained from a relationship with a classification-granted patent gazette, so that the same effect can be obtained even if the non-classified patent gazette set 100 is divided, and the processing time is reduced. Is done.

It is a block diagram of the document classification provision apparatus which concerns on Example 1 of this invention. It is explanatory drawing of the distance between documents. It is a patent gazette two-dimensional coordinate mapping figure at the time of weighting 10 times with respect to the value of the first IPC at the time of document vector creation. It is a patent gazette two-dimensional coordinate mapping figure in case weighting is not carried out to the first IPC at the time of document vector creation. It is explanatory drawing of the selection object area | region in a patent gazette two-dimensional coordinate mapping figure. It is explanatory drawing of the area | region in a patent gazette two-dimensional coordinate mapping figure. It is a figure which shows the classification | category assignment | providing accuracy in the area | region "A" shown in FIG. It is a patent publication three-dimensional coordinate mapping figure at the time of adding the number of patent publications as a z-axis with respect to a patent publication two-dimensional coordinate mapping figure. It is the flowchart which showed the operation | movement in the non-classification | assignment subset selection part which concerns on Example 1 of this invention. It is a block diagram of the document classification provision apparatus which concerns on Example 2 of this invention. It is the flowchart which showed the operation | movement in the classification removal part which concerns on the same Example. It is a block diagram of the document classification provision apparatus which concerns on Example 3 of this invention. It is explanatory drawing of the similarity between 3 characters.

Explanation of symbols

DESCRIPTION OF SYMBOLS 11 Document input part 12 Feature-value extraction part 13 Document feature-value vector generation part 14 Main feature-value dictionary 15 Inter-document distance calculation part 16 Dimensional coordinate mapping process part 17 Unclassified subset selection part 18 Classification grant object patent publication subset 20 Classification-added document input unit 21 Filtering feature amount extraction unit 22 Classification removal unit 91 Unclassified document set division unit 92 Divided non-granted patent gazette set 100 Non-classified patent gazette set 101 Classification-granted patent gazette set

Claims (11)

Document classification assignment that assigns the predetermined classification category to a non-classification document that is a document that has not been assigned a classification category based on a classification grant document that is a document to which a predetermined classification category has been assigned In the device
A document input unit for inputting a plurality of documents including the classified documents and the unclassified documents;
A feature amount extraction unit that extracts a feature amount from each document input from the document input unit;
A main feature dictionary storage unit for storing a rule for extracting a main feature from the feature;
A main feature amount is extracted from the feature amounts obtained from the feature amount extraction unit with reference to the rules stored in the main feature amount dictionary storage unit, and a weight greater than 1 is superimposed on the value of the main feature amount A document feature quantity vector generation unit for generating a feature quantity vector having the feature quantity as an element after
An inter-document distance calculation unit that calculates a distance between documents using the feature vector generated by the document feature vector generation unit;
A two-dimensional coordinate mapping processing unit that arranges coordinate points corresponding to each document on two-dimensional coordinates so that the relationship between the distances between documents obtained by the inter-document distance calculation unit appears.
From the two-dimensional coordinates, an unclassified document for selecting a set of coordinate points corresponding to the unclassified document having the main feature amount and to which the predetermined classification category should be assigned. A document classification assigning device comprising: an assignment subset selection unit. In Claim 1, the filtering feature-value extraction part which extracts the feature-value for filtering used for the similarity determination of a document from the classification provision document input from the said document input part,
For each of the predetermined classification categories, the similarity between the shape of the filtering feature amount frequency distribution of the entire classification-giving document and the shape of the filtering feature amount frequency distribution of each of the classification-giving documents Based on the determined similarity, the frequency distribution of the filtering feature amount of the predetermined classification-added document is similar to the shape of the frequency distribution of filtering feature amount of the entire classification-added document. A filtering feature amount having the highest frequency in the frequency distribution of filtering feature amounts of the entire classification-giving document and the predetermined classification-giving document. If the filtering feature quantity with the highest frequency in the frequency distribution of filtering feature quantities is not the same, it is given to the predetermined classification grant document. Anda classification removing unit for removing the classification category,
An apparatus for assigning document classification, wherein the document is input to the two-dimensional coordinate mapping processing unit in order to arrange coordinate points corresponding to the document obtained by the classification removing unit on the two-dimensional coordinate.   3. The document classification assigning apparatus according to claim 2, wherein the classification removing unit uses a chi-square test process for determining the similarity of the shape of the frequency distribution of the feature quantity for filtering.   4. The unclassified subset selection unit according to claim 1, wherein the unclassified subset selection unit sets a circular area including a coordinate point corresponding to the classified document on the two-dimensional coordinates, and is included in the circular area. And a set of coordinate points corresponding to a non-classified document having the main feature amount is selected.   5. The document according to claim 1, wherein the plurality of input documents is a patent gazette, and the feature amount of the extracted document and the feature amount for filtering are IPC (International Patent Classification). , FI, FT (File Forming Term) or a word in a patent gazette, the classification category is a technical classification category, and the rule stored in the main document feature dictionary storage unit is the first A document classification assigning apparatus, characterized by extracting an IPC, an FI similar in content to the first IPC, an FT related to the FI, or a word in a patent publication. Performed by the document classification assigning device for assigning the predetermined classification category to the unclassified document without the classification category based on the classification grant document that is a document to which the predetermined classification category is assigned. In the document classification assignment method,
A document input step of inputting a plurality of documents including the classified document and the unclassified document;
A feature amount extraction step of extracting a feature amount from each document input from the document input step;
A main feature dictionary storing step for storing a rule for extracting a main feature from the feature in the main feature dictionary storage;
A main feature amount is extracted from the feature amounts obtained in the feature amount extraction step with reference to the main feature amount dictionary storage unit, and the feature amount after a weight greater than 1 is superimposed on the main feature amount A document feature vector generation step for creating a feature vector as an element;
An inter-document distance calculation step of calculating a distance between documents using the feature amount vector generated in the document feature amount vector generation step;
A two-dimensional coordinate mapping processing step of arranging coordinate points corresponding to each document on the two-dimensional coordinates so that the relationship between the distances of the documents obtained in the inter-document distance calculation step appears.
A non-classified portion for selecting a coordinate point corresponding to a non-classified document to which the predetermined classification category is to be assigned and which is the non-classified document having the main feature amount from the two-dimensional coordinates A document classification assigning method comprising: a set selection step. The filtering feature value extracting step according to claim 6, wherein a filtering feature value for use in document similarity determination is extracted from the classified document input from the document input step;
For each of the predetermined classification categories, determine the degree of similarity between the filtering feature amount frequency distribution of the entire classification-giving document and the filtering feature amount frequency distribution of each of the classification-giving documents, Based on the determined similarity, it is determined that the frequency distribution of the filtering feature value possessed by the predetermined classification grant document is not similar to the shape of the frequency distribution of the filtering feature amount possessed by the entire classification grant document. And the filtering feature amount having the highest frequency in the frequency distribution of the filtering feature amount possessed by the entire classification-giving document and the filtering feature possessed by the predetermined classification-giving document. When the feature quantity for filtering with the highest frequency in the frequency distribution of quantities is not the same, the classification category assigned to the predetermined classification grant document is Anda classification removal step that support,
In the two-dimensional coordinate mapping processing step,
A document classification assigning method, wherein coordinate points corresponding to a document obtained in the classification removal step are arranged on the two-dimensional coordinates.   8. The document classification assigning method according to claim 7, wherein in the classification removal step, a chi-square test process is used for determining the similarity of the shape of the frequency distribution of the filtering feature quantity.   9. The non-classified subset selection step according to any one of claims 6 to 8, wherein a circular area including a coordinate point corresponding to the classified document is set on the two-dimensional coordinates and is included in the circular area. And a document classification assigning method, wherein a set of coordinate points corresponding to an unclassified document having the main feature amount is selected.   10. The input document according to claim 6, wherein the plurality of input documents are patent gazettes, and the feature amount of the extracted document and the feature amount for filtering are IPC (International Patent Classification). , FI, FT (File Forming Term) or a word in a patent gazette, the classification category is a technical classification category, and the rule stored in the main document feature dictionary storage unit is the first A document classification assigning method, characterized by extracting an IPC, an FI similar in content to the first IPC, an FT related to the FI, or a word in a patent gazette.   A program for causing a computer to execute the method according to any one of claims 6 to 10.

Images