JP2005339412A - Patent map generation method and program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method for automatically or semi-automatically generating a patent map by a computer or the like to solve the problem of quantitative limitation and high cost, resulting from manual formation of the patent map only by patent experts of each technical object field. <P>SOLUTION: A conception base 204 that is a knowledge base for words contained in a patent specification document group 201 that is to be an object of patent map generation is generated by statistically processing the ways to use words in the patent specification document group 201, and vector values 205 for individual patent specifications contained in the patent specification document group 201 are calculated by the use of the concept base 204. By clustering the patent specification group 201 based on the vector values 205, the positions of the individual patents in the patent specification document group 201 are clarified, based on which clusters patents shown by the individual patent specifications of the patent specification document group 201 belong to, and the positions are visualized, whereby the patent map 303 is generated. <P>COPYRIGHT: (C)2006,JPO&NCIPI

Description

  In the present invention, when a plurality of patents related to a specific object are given, the mutual relations of the patents included in those patent groups are classified based on the contents, and the classified patent groups are classified according to the classification status. The present invention relates to a patent map generation technique for displaying in a visually easy-to-understand manner, and more particularly to a micromap generation technique.

  The patent map includes a macro map that visualizes application trends based on a group of patents related to a specific technical field, and a relationship between patents included in the group of patents based on one or more viewpoints regarding a group of patents related to a specific object And a micromap that visualizes how each patent is positioned in the subject invention.

  The present invention relates to a micromap. Conventionally, the creation of a micromap requires technical knowledge about the technical object and knowledge about the format of the patent document, and patent experts in the technical field to which the technical object belongs. Was created manually.

  However, in such background technology, since the number of patent experts for each technical field to which the technical object belongs is small, the number of patent maps that can be created in a certain period is limited, and enormous costs are required. This is a problem. Despite the increasing importance of patent maps with the increasing awareness of companies and individuals about intellectual property, the creation of patent maps depends only on the skills of a few specialists. The national patent strategy is also a big problem.

  An object of the present invention is to automatically or semi-automatically use a computer or the like in order to solve the problem of quantitative limitations and expensive costs caused by a patent expert in the technical subject field creating a patent map only by hand. It is intended to provide a method for generating a patent map.

  The present invention generates a “concept base”, which is a knowledge base related to a word included in a patent specification document group that is a target of patent map generation, by statistically processing how the word is used in the patent specification document group. And calculating a vector value for each patent specification document included in the patent specification document group using the concept base, and clustering the patent specification document group based on the vector value, thereby obtaining the patent specification document Based on which cluster the patents represented by the individual patent specifications of the group belong, the patent map is generated by clarifying the position of the individual patent in the patent specification document group and visualizing the position.

Hereinafter, means for solving the problems in the present invention will be described more specifically.
A collection of patent specification documents relating to a particular subject is stored in a storage device. Based on a given viewpoint for generating a patent map, a portion related to the viewpoint is extracted from each patent specification document of a set of patent documents (referred to as a patent specification document group) stored in the storage device. . The part extracted from the patent specification document shall be representative of the patent specification document in patent map generation. The extracted portion of the patent specification document is subjected to a splitting process and stored as a word string in the storage device. By applying this process to all patent specification documents included in the patent specification document group, a group of word strings corresponding to the extracted patent specification document is stored in the storage device.

  Statistical processing is performed on the word string group based on the appearance frequency of the word, the relationship between the word and the word appearance distance, etc., and all the different words included in the word string group are assigned to individual words. The corresponding vector value is calculated and stored in the storage device. By this processing, for the word appearing in the extracted part of the patent specification document, the word and the vector value corresponding to the word are obtained as a set. A set composed of a set of this word and a vector corresponding to the word is called a concept base. By retrieving a word from the concept base, a vector value corresponding to the word can be obtained.

  After creating the concept base, for each extracted patent specification document part in the group of extracted patent specification document parts, the word vector from the concept base with respect to the individual words contained in the corresponding word string The values are obtained and the vectors corresponding to all the words over the word string are summed, and this is used as the patent vector value of the patent represented by the patent specification document. That is, according to the processing so far, the corresponding patent vector value is assigned to the patent represented by each patent specification document included in the given patent specification document group.

  After patent vector values are obtained for patents represented by all patent specification documents stored in the storage device, the patent groups represented by the patent specification document groups are clustered based on the patent vector values. By the clustering process, the patent group is classified into one or more clusters.

  When multiple viewpoints for classification are given, the above-described processing from extracting a part related to the viewpoint from the patent specification document to the clustering process is repeated for each viewpoint, for each patent. Classification in terms of

A patent map is obtained by arranging a given group of patents on a two-dimensional plane such as a paper surface or a screen based on the obtained classification in each viewpoint.
In the patent map generation method of the present invention, a portion related to the viewpoint of creating a patent map is extracted from the patent specification document of a given patent, the extracted portion of the document is divided into word strings, and included in the word string By applying statistical processing to each word, a vector value corresponding to each word is obtained. The resulting vector value reflects how the word is used in the group of parts of the patent specification document, which represents the meaning of the word in context in the group of parts of the patent specification document as knowledge. You can also say.

  With respect to the parts of the patent specification document included in the group of parts of the patent specification document, a vector value corresponding to the word is obtained from the concept base for each word in the word string corresponding to the part of the patent specification. The vector values obtained in this way are summed over all the words included in the word string to obtain a patent vector value corresponding to the patent represented by the patent specification document. Since a document is composed of a set of words and is considered to be the total knowledge of words, the patent vector value obtained by this processing expresses the characteristics of the patent based on the viewpoint for generating the patent map. Will be.

  Since patent vector values express the characteristics of patents in terms of patent map generation, clustering patent groups based on patent vector values can provide classification in terms of patent map generation. .

  When a plurality of viewpoints are given, the above is repeated for each viewpoint, whereby a classification for each viewpoint can be obtained for a given patent group. For example, when two viewpoints are given, individual patents included in the patent group belong to two clusters having different viewpoints. It can be said that it can be classified into two dimensions. Similarly, when three viewpoints are given, individual patents included in the patent group belong to three clusters having different viewpoints. That is, it can be said that it can be classified into three dimensions.

  Based on the classification obtained above, finally, the patent group is visually represented on a paper or a screen. Since the paper and the screen are two-dimensional planes, if the viewpoint for generating a patent map is three or more, the paper and the screen are reduced to two dimensions by some method.

  According to the method of the present invention and the program based thereon, a set of patent specification documents relating to a specific object is given, and one or more viewpoints for generating a patent map are given. It is possible to generate a patent map in which the patents are arranged on a two-dimensional plane based on the mutual relationship between the patents represented by the individual patent specification documents. Conventional patent map generation is performed manually by a patent expert who is familiar with the technical field in which the object is included, but according to the present invention, a patent map can be generated automatically or semi-automatically by a computer or the like.

Hereinafter, the best mode for carrying out the present invention will be described in detail with reference to the drawings.
FIG. 1 shows a processing procedure and a data flow of an embodiment of the present invention.

  A set 301 of patent specification documents relating to a given specific object is stored in a storage device and becomes a patent specification document group 201. Here, the present invention essentially does not prescribe how to search for a patent specification document related to a specific object. However, a human being gives a logical expression using keywords and keywords suitable for a patent search system. A method such as obtaining a patent specification document group is conceivable.

  At the same time, a viewpoint 302 for creating a patent map is designated from the outside. Based on the specified viewpoint, a specification of the portion of the patent specification document corresponding to the viewpoint is generated at 101. For example, when “problem to be solved” is given as a viewpoint, it is designated to extract the “problem to be solved by the invention” from the patent specification document. The specific method for making the designation is not essentially defined in the present invention. However, when a table of viewpoints and parts corresponding to the viewpoints is incorporated in advance, and the viewpoint is given. In addition, a method of searching for a description about a corresponding part corresponding to the viewpoint from the table can be considered.

  When a portion corresponding to a given viewpoint is specified, extraction of a related portion of the patent specification document is performed at 102 based on the specification from the individual patent specification documents of the patent specification document group 201 stored in the storage device. The extracted patent specification part group 202 is stored in the storage device.

  The morpheme analysis process 103 is performed on the individual patent specification document parts included in the group of extracted patent specification document parts 202. In the morpheme analysis process, a sentence included in an input document part is decomposed into word strings. That is, by dividing a sentence into words, a Japanese sentence is converted into a string of words included in the sentence. All sentences included in the extracted patent specification document part group 202 are converted into word strings by the morphological analysis processing 103 and stored as a word string group 203 corresponding to the extracted patent specification document parts. Stored in the device.

  A neighborhood co-occurrence matrix is generated by the neighborhood co-occurrence matrix generation 104 from the word string group 203 corresponding to the extracted part of the patent specification document, the dimension compression 105 is performed by singular value decomposition, and the concept base 204 is generated. Is done. In the following, the concept-based generation will be described in more detail with reference to FIG.

  As shown in the word string group 401 in FIG. 2, the word string group 203 corresponding to the extracted patent specification document part in FIG. The included words form a sequence of words in the order of appearance. The group of word strings is taken as an input, and a neighborhood co-occurrence matrix is generated in neighborhood co-occurrence matrix generation 104. Neighboring co-occurrence matrix means that a word (we write this as Wi) and a word (we write this as Wj) occur simultaneously within a certain number of words in all word strings to be processed. This is a matrix that has the values calculated from the frequency to be used as the elements of the row corresponding to Wi and the column corresponding to Wj. The method for generating the neighborhood co-occurrence matrix is not essentially defined in the present invention. However, when the number of words included in the word string is n, n rows n in which all elements are 0. First, a matrix of columns is prepared, and for each word Wj that appears within a certain number of words before and after the word Wi while scanning a group of word sequences in sequence, a row corresponding to Wi, corresponding to Wj By adding a certain number to the elements of the sequence to be generated, etc., it can be generated with an amount of calculation in an order proportional to the number of words included in the word sequence group. Also, what value is used as a value calculated from the frequency of co-occurrence in the vicinity is not essentially defined by the present invention. It is conceivable to use the frequency itself or to use a value obtained by inputting the frequency into a monotonically increasing function. In addition, the neighborhood co-occurrence matrix is not necessarily a square matrix having the same number of rows and columns. The word vector values associated with individual words are obtained as individual rows in the matrix, so the number of rows must be the same as the number of words required as a concept base, but the number of columns must be equal to this. Absent.

When the neighborhood co-occurrence matrix 402 is generated, the neighborhood co-occurrence matrix is subjected to singular value decomposition by the dimension compression processing 105 based on singular value decomposition, and then dimension compression is performed based on the result.
Singular value decomposition means that when there is a matrix A,
A = L × λ × R
Is a process for obtaining matrices L, λ, and R such that However, λ is a diagonal matrix, that is, a matrix in which all elements other than the diagonal elements are zero. In general, the diagonal elements of λ are determined so that the values in the upper row are larger. When the rank of the matrix A is r, the matrix L has a size of n rows and r columns, the matrix λ has a size of r rows and r columns, and the matrix R has a size of r rows and m columns. .

  After performing the singular value decomposition, only some columns are selected and extracted from the left matrix L obtained by the decomposition to obtain a new matrix 404. In the neighborhood co-occurrence matrix 402, when the row corresponding to a certain word Wi is the i-th row, the i-th row of the new matrix 404 is a word vector value associated with the word. The concept base 204 is a combination of this word and the word vector value associated with the word stored in the storage device for all words included in the word string group. Which column is selected as the word vector value from the left matrix is not essentially defined in the present invention, but in general, a certain number of columns are selected from the left column. Specifically, the neighborhood co-occurrence matrix is about 10000 rows × 3000 columns, and 100 to 200 columns are selected from the left side of the left side matrix. Of course, these numbers are merely examples in the present embodiment, and are not essentially defined in the present invention.

  In the dimension compression 105 based on singular value decomposition, the method of performing singular value decomposition is not essentially defined in the present invention. In general, a square matrix B having the matrix A at the upper left and the transpose of the matrix A at the lower right and the other elements being 0 is generated, the eigenvalue of the square matrix B is obtained, and the eigenlocation is determined as A Singular value decomposition can be performed by making it correspond to singular values. In addition, in the neighborhood co-occurrence matrix A, when the number of elements that are 0 is relatively high, that is, when A is a sparse matrix, a ranchos that can perform singular value decomposition with a small memory requirement and high speed. Laws are used. Furthermore, since it is only necessary to finally obtain the element values in the required columns of the left side matrix, after obtaining all the elements of the left side matrix, the selected columns are extracted and used as word vector values. It is not always necessary to perform a two-stage process. Some processing methods of singular value decomposition obtain the column values of the left matrix sequentially, so when using this method, the singular value decomposition is performed when all the required column values are obtained. The value decomposition process may be aborted.

  After the concept base is generated, a patent vector value generation 106 for the patent represented by each patent specification document is performed from the group of word strings 203 corresponding to the extracted part of the patent specification document. A patent including a part of the extracted patent specification document by using a word string of each extracted part of the patent specification document included in the word string group 203 corresponding to the extracted part of the patent specification document The patent vector value of the patent represented by the specification document will be calculated. For each word included in the word string of each extracted part of the patent specification document, a word vector value associated with the word is obtained on a concept basis, and the obtained word vector value is All the words included in the word string of the extracted part of the patent specification document are added to obtain a combined vector value. This combined vector value is used as the patent vector value of the patent represented by the patent specification document including the extracted portion of the patent specification document. In this embodiment, the synthesized vector is obtained by simply adding the vector values of the words. However, this is not essentially defined in the present invention, and each word vector value depends on the characteristics of the word. A method of summing up some weight is also conceivable.

  As described above, for each patent of the patent group represented by the patent specification document group 201, the patent vector value associated with the patent is obtained. Patents in which all the patents of the patent group represented by the patent specification document group 201 are stored in the storage device as a set of a code (for example, a patent number) that can uniquely designate the patent and a patent vector value associated with the patent. A vector value group 205 is obtained.

  The patent vector value group 205 is classified by the clustering 107 based on the patent vector value, and is stored in the storage device as a set of a code for designating a patent and a code for the classification to which the patent belongs. Assuming that all the patents represented by the patent specification document group 201 are classified and stored, the correspondence information 206 between the patent and the classification is obtained. Here, the clustering method is not essentially defined in the present invention. A hierarchical clustering method may be used, or a non-hierarchical clustering method may be used. In hierarchical clustering, a cluster hierarchy diagram can be obtained by clustering, and it is a problem to classify based on which hierarchy on this hierarchy diagram, but there is a similarity between clusters found in clustering The highest layer is selected from layers that do not become smaller than a preset threshold value. However, this cluster hierarchy selection criterion is not essentially defined in the present invention, and a method of determining a cluster hierarchy to be used for classification based on other information given from the outside is also conceivable. Further, in the non-hierarchical clustering, it is necessary to give the number of clusters, but the method for determining the given number of clusters is not essentially defined in the present invention.

  Based on the obtained correspondence information 206 between the patent and the classification, the patent is placed on the two-dimensional plane, and based on the placement, the target patent is placed on the paper or the screen on the two-dimensional plane. As a result, a patent map 303 displayed as is obtained. In the present embodiment, the case where there is one viewpoint is described as an example, so the classification axis is one-dimensional and the arrangement on a two-dimensional plane is easy.

  However, the number of viewpoints given is not essentially defined in the present invention. When two or more viewpoints are given, the specification specification document portion specification 101 to clustering 107 related to the viewpoint is performed for each viewpoint, and the correspondence information 206 between the patent and the classification based on the clustering is also displayed for each viewpoint. To get to. Based on the patent-to-classification correspondence information 206 obtained for each viewpoint, the patents are arranged on a two-dimensional plane. When there are two viewpoints, it can be easily arranged as a table. When there are three or more viewpoints, a process of projecting on a two-dimensional plane is required, but the processing method is not essentially defined in the present invention. In the case where there are three viewpoints, a method of arranging as a table based on two of the three viewpoints, and expressing the classification based on the remaining one viewpoint as the size of a character or a point representing a patent, or the like can be considered. When there are four viewpoints, in addition to the projection when there are three viewpoints, a method of expressing the classification of another viewpoint as a color or the like can be considered.

It is a procedure figure of the processing method in one Example of this invention. It is a procedure figure of the concept base production | generation method in this invention.

Explanation of symbols

101 Patent specification document part specification related to viewpoint 102 Extraction of patent specification document part based on extraction specification 103 Morphological analysis processing 104 Neighborhood co-occurrence matrix generation 105 Dimensional compression by singular value decomposition 106 Patent vector value generation 107 Clustering 108 Two-dimensional plane Arrangement in 201 Patent specification document group 202 Extracted patent specification document part group 203 Word string group corresponding to extracted patent specification document part 204 Concept base 205 Patent vector value group 206 Correspondence information of patent and classification 301 Patent Specification Document for Creating Patent Map 302 Perspective for Creating Patent Map 303 Patent Map

Claims (6)

A computer program for causing a computer to execute a process of generating a patent map by classifying the mutual relations of patents included in a group of patents based on the contents when a plurality of patents relating to a specific object are given Because
For patent specification documents stored in the storage device,
A process of extracting a specific part from individual patent specification documents of the patent specification document group;
A process of dividing the specific part into words to form a word string;
By applying statistical processing to word strings in all patent specification documents of the patent specification document group, for all words included in the word string, individual words and vector values associated with the individual words A process based on a concept stored in a storage device;
For individual patent specification documents of the patent specification document group stored in the storage device,
A process of obtaining a word set by dividing a specific part extracted from the patent specification document for each word;
For each word included in the word set, obtaining a vector value associated with the word from the concept base;
A process of calculating the sum of vector values obtained for all words included in the word set;
Calculating a vector value that is the sum of the vectors as a vector value of a patent represented by the patent specification document, and storing the calculated value in a storage device in association with the patent;
Processing for clustering the patent group represented by the patent specification document group stored in the storage device based on the patent vector values associated with the individual patents of the patent group;
A process of performing classification for each patent belonging to the same cluster;
A computer program that executes. A computer program according to claim 1,
For a word string group for a set of parts of a patent specification document obtained by extracting a specific part from a group of patent specification documents stored in a storage device,
Counting different words in the word string,
A process of creating a matrix in which the individual words of the counted words correspond to rows and the individual words also correspond to columns;
When the word string group is sequentially scanned and a word (hereinafter referred to as the word Wi) and a word (hereinafter referred to as the word Wj) appear nearby in the word string, the line corresponding to the word Wi of the matrix Incrementing the number of elements in the column corresponding to the word Wj by a positive number is repeated for all the word strings included in the word string group to obtain the value of each element of the matrix (hereinafter referred to as A matrix of element values is called a neighborhood co-occurrence matrix),
Processing for singular value decomposition of the neighborhood co-occurrence matrix;
A process of selecting several columns from the columns of the right matrix calculated by the singular value decomposition;
A process of creating a concept base by associating a word value with a vector value that also has a numerical value of an element of the selected column as an element in a row corresponding to the word in the right side matrix;
A computer program that executes. A computer program according to claim 1,
A computer program for executing a process of arranging and displaying a patent, which is a result of executing classification for each patent belonging to the same cluster, in the vicinity. A method for generating a patent map by classifying a mutual relationship of patents included in a group of patents based on contents when a plurality of patents related to a specific object are given,
For patent specification documents stored in the storage device,
Extracting specific parts from individual patent specification documents of the patent specification document group;
The specific part is divided into words to form a word string,
By applying statistical processing to word strings in all patent specification documents of the patent specification document group, for all words included in the word string, individual words and vector values associated with the individual words Is stored in a storage device as a concept base,
For individual patent specification documents of the patent specification document group stored in the storage device,
A process of obtaining a word set by dividing a specific part extracted from the patent specification document for each word;
For each word included in the word set, obtain a vector value associated with the word from the concept base;
A process of calculating the sum of vector values obtained for all words included in the word set;
A vector value that is the sum of the vectors is calculated as a vector value of a patent represented by the patent specification document, and stored in a storage device in association with the patent;
Clustering is performed on the patent group represented by the patent specification document group stored in the storage device based on the patent vector values associated with the individual patents of the patent group,
Perform classification for each patent belonging to the same cluster,
A method for generating a patent map, comprising a process. The method of claim 4, comprising:
For a word string group for a set of parts of a patent specification document obtained by extracting a specific part from a group of patent specification documents stored in a storage device,
Count different words in a word string,
Create a matrix in which the individual words of the counted words correspond to rows and the individual words also correspond to columns,
When a sequence of word strings is scanned sequentially and a word (hereinafter referred to as word Wi) and a word (hereinafter referred to as word Wj) appear nearby in the word string, the line corresponding to word Wi in the matrix Increasing the number of elements in the column corresponding to the word Wj by a positive number is repeated for all the word strings included in the word string group, and the values of the elements of the matrix are obtained (hereinafter referred to as found A matrix of element values is called a neighborhood co-occurrence matrix),
Singular value decomposition of the neighborhood co-occurrence matrix,
Select some columns from the columns of the right matrix calculated by the singular value decomposition,
Creating a concept base by associating a word value with a vector value having the numerical value of the element of the selected column as an element in a row corresponding to the word in the right-side matrix;
A method for generating a patent map, comprising a process. The method of claim 4, comprising:
A method for generating a patent map, comprising a step of arranging a patent that is a result of executing classification for each patent belonging to the same cluster and displaying it on the screen in the vicinity.

Images