JP2000123041A - Similarity judging method, document retrieving device, document classifying device, storage medium stored with document retrieval program and storage medium stored with document classification program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a similarity judging method, a document retrieving device and a document classifying device capable of precisely judging similarity between document element having plural subjects or sub-titles, judging the similarity using the feature of the document elements and precisely judging the similarity by solving the incompleteness of morphemic analysis, and to provide a storage medium in which a document retrieval program is stored and a storage medium in which a document classification program is stored. SOLUTION: The word used in document element composed of word string, the Boolean operator join of the word string, sentence, document, or document set is extracted from the document element (S1), the criticality is imparted to the extracted word (S2), and the relevancy degree is imparted to the extracted two arbitrary words (S3). Then, the subjects of respective document elements are expressed by graphics prepared with the criticality of the word as the weight of a node and with the relevancy degree between those words as the weight of the link (S4), and the similarity between the document elements is judged based on the matching degree between graphics expressing the subjects (S5).

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a similarity determination method, a document search device, a document classification device, a storage medium storing a document search program, and a storage medium storing a document classification program. , A Boolean operator combination of word strings, or a similarity determination method, a document search device, a document classification device, and a document search program for appropriately determining the similarity between document elements composed of documents or document sets. And a storage medium storing a document classification program.

[0002]

2. Description of the Related Art A conventional document search apparatus determines a search result by determining the similarity between a search key and a search target document by the following method. Here, the search key may be not only a word string or a Boolean operator combination of the word strings but also a document or a document set.

[0003] First, words used in each document to be searched are extracted using a technique called morphological analysis, and based on appearance frequency information and the like, each word is associated with the subject (content) of the word. Is given a degree of importance indicating the strength of Similarly, we assign importance to the words in the search key entered by the user, check how important each word is in each document, and change the Boolean operator in the search key. Appropriate processing is performed to calculate the similarity.
Finally, the documents to be searched are sorted in descending order of the similarity thus obtained, and the top n items are set as the search results. Here, n is a positive constant.

For example, “information or a document (especially a document)
Suppose you want to search for documents related to "search." The user
Specify ((information ^- 0.5 or document ^- 0.8) and search ^- 1.0) as the search key. First, the document search apparatus first determines the importance of a word in a search key ((information, 0.5) (document, 0.8)
(Search, 1.0) is obtained (in this example, the importance is explicitly described in the search key, but these importance may be automatically determined from word appearance frequency information or the like). .

Next, the importance of the words "information", "document", and "search" used in the search key in each document to be searched is determined based on the frequency of appearance. Suppose these were the following values: Calculate the product of the importance of the word in the search key and the importance of the word in the search target document, and if or is used in the search key, add the values on both sides, If and is used, the smaller of the values on both sides is taken.

With this method, the similarity between a search key and each document can be obtained as follows. Similarity of document a = min ((0.5 * 0.4 + 0.8 * 0.6), 1.0 * 0.9) =
min (0.68,0.9) = 0.68 Similarity of document b = min ((0.5 * 0.4 + 0.8 * 0.1), 1.0 * 0.0) =
min (0.26,0.0) = 0.0 Similarity of document c = min ((0.5 * 0.3 + 0.8 * 0.8), 1.0 * 1.0) =
min (0.79,1.0) = 0.79 Here, min (x, y) returns the smaller value of x and y.

In this way, the similarity between the search key and the search target document is calculated, the search target documents are sorted in descending order of this value, and the top n (= 2) items are set as the search results. Therefore, the search result in this case is document c (similarity 0.79) and document a (similarity 0.68).

Further, the conventional document classifying apparatus determines the similarity between all two documents in a set of documents to be classified by determining the similarity between the documents.
Classifies documents. First, words are extracted from each document of a set of documents to be classified, and the words are given appropriate importance. Next, based on this importance,
The similarity between all two documents in the set of documents to be classified is determined in the same manner as the method described in the document search apparatus. Next, based on the similarity between the documents, the documents are classified by sequentially combining documents having a large similarity. This technique is called clustering.

[0009]

In order to construct a more accurate document search device and document classification device, it is necessary to appropriately determine the similarity between document elements. Here, the document element is a word string or a Boolean operator combination of words,
Alternatively, it is a document or a document set. However, the conventional similarity determination method has the following problems.

[0010] 1. The similarity between document elements with multiple themes and subtitles cannot be determined with high accuracy: When a document element is a word string or an abstract, it is considered that the document element has one subject. When targeted, the document will have multiple subjects and subtitles. Therefore, the similarity is not properly calculated when the entire text of such a document is targeted.

For example, in a document search operation, it is assumed that a user specifies a search key of "information search and robot" when the user wants to search for a document related to "robot performing information search". However, this search key gives a high degree of similarity to documents having two themes, "information retrieval system" and "industrial robot". As described above, when the document element has a plurality of themes and subtitles, there is a problem that the similarity cannot be accurately determined.

2. It is not possible to judge similarity using features of document elements: between words used in a sentence,
There are characteristics such as dependency relationship. Documents also have features such as paragraphs. However, the conventional similarity determination method only extracts words, assigns importance to those words, and determines similarity based on the words. Therefore, these features cannot be used, and similarity determination cannot be performed. Cannot be determined with high accuracy.

3. Imperfect morphological analysis: In morphological analysis used when extracting words from a document, it is necessary to recognize which partial character string is to be a word, and therefore, it is necessary to register the word in a dictionary in advance There is. However, when the update speed of information is high, it is impossible to register all words in a dictionary in advance, and when such information is targeted, the analysis fails when extracting words. Is inevitable. For example, if the dictionary contains only the words "inter" and "net",
The word “internet” is not extracted, and this word is extracted as two words “inter” and “net”. As described above, there is a problem that the similarity cannot be determined with high accuracy because the word extraction fails.

The present invention has been made in view of the above points, and it is possible to accurately determine the similarity between document elements having a plurality of subjects and subtitles and determine the similarity using the features of the document elements. A similarity determination method, a document search device, a document classification device, and a storage medium and a document storing a document search program that can resolve incompleteness of morphological analysis and accurately determine the similarity between document elements It is an object to provide a storage medium storing a classification program.

[0015]

FIG. 1 is a diagram for explaining the principle of the present invention. The present invention (Claim 1) provides a similarity determination method for appropriately determining the similarity between document elements, wherein a word string, a Boolean operator combination of word strings, a sentence, a document, or a document set is used. From the composed document elements, words used in the document elements are extracted (step 1), and the extracted words are assigned importance (step 2), and each of the extracted arbitrary 2 words is assigned. Relevance is given between words (step 3), and the subject of each document element is expressed by a graph in which the importance of the word is used as the weight of the node and the relevance between the words is used as the weight of the link (step 4). ), The similarity between the document elements is determined based on the degree of agreement between the graphs representing the themes (step 5).

According to the present invention (claim 2), when calculating the degree of coincidence between graphs for expressing a subject, the number of similar nodes (nodes containing the same word) in both graphs is large. The greater the number, the greater the degree of coincidence between the graphs, and if a node in one graph has a large weight, if a similar node in the other graph has a large weight, The greater the number of similar links (links containing the same word in the nodes at both ends of the link) in both graphs, the greater the degree of matching between the two graphs. If the degree of agreement between the graphs is set to a large value, and a link in one graph is heavily weighted, the more the similar link in the other graph is weighted, the more One between graphs To the degree of a large value, to calculate the degree of match between the graph representing the subject.

According to the present invention (claim 3), when calculating the degree of coincidence between graphs for expressing a subject, each graph is represented by the strength of a word set used in the graph. Is divided into subgraphs based on whether the subgraphs are related to each other, and if each subgraph has no link between any nodes in the subgraph, a link having a small weight is generated in the subgraph. Then rejoin the subgraphs, add the links you just created to the subgraphs,
Return to the graph before the division and calculate the degree of agreement between the graphs expressing the subject.

According to the present invention (claim 4), when calculating the degree of coincidence between graphs for expressing the subject, each graph is determined by the strength of the word set used in the graph. Is divided into subgraphs based on whether the subgraphs are related to each other. If there is no link between any nodes between the subgraphs in each subgraph, a link having a small weight is assigned to the subgraph. Is generated, and the degree of matching is calculated for each subgraph.

According to the present invention (claim 5), when calculating the degree of coincidence between graphs for expressing a subject, each graph is represented by the strength of a word set used in the graph. Is divided into subgraphs based on whether the subgraphs are related to each other, and if each subgraph has no link between any nodes in the subgraph, a link having a small weight is generated in the subgraph. Then, the degree of coincidence is calculated for each of the subgraphs, and the sum of the degrees of coincidence calculated for each of the subgraphs is calculated, thereby calculating the degree of coincidence between the graphs expressing the subject.

FIG. 2 is a block diagram showing the principle of the document search apparatus according to the present invention. The present invention (claim 6) is a document search apparatus for searching for a document based on a search request from a user, the search interface means 410 for analyzing a search request from a user and extracting a search key, A search key subject graph creation unit 420 that generates a graph representing the subject of the search key from the key, and a word information management unit 430 that acquires a set of document IDs of documents in which a specified word appears.
When a document ID is specified, a document corresponding to the document ID is obtained from the search target document storage unit 441, and a search target document subject graph creation unit 440 that creates a graph expressing the subject of the document, A graph expressing the subject of a key and a graph expressing the subject of a search target document are input, and a similarity determination unit 450 for determining how similar they are, a search interface unit 410, a search key subject graph creation unit 420, a word information management unit 430, a search subject document subject graph creation unit 440, and a search control unit 460 that controls the similarity determination unit 450.

According to the present invention (claim 7), in the search key subject graph creating means 420, a word string, a Boolean operator combination of a word string, or a document element composed of a sentence, a document, or a document set is used. Word extracting means for extracting words used in the document element; importance assigning means for assigning importance to each of the extracted words; and relevance between any two extracted words. The search subject document subject graph creating unit 440 includes a relevance assigning unit that assigns the relevance, and the search target document subject graph creating unit 440 uses a graph in which the importance of the word is set as the weight of the node and the relevance between the words is set as the weight of the link. Including subject expressing means for expressing the subject,
The similarity determination unit 450 includes a unit that determines the similarity between document elements based on the degree of coincidence between graphs representing subjects.

According to the present invention (claim 8), the similarity determination means 4
At 50, the greater the number of similar nodes (nodes containing the same word) in both graphs, the greater the degree of matching between the graphs, the greater the weight of the nodes in one graph If so, the greater the weight of similar nodes in the other graph, the greater the degree of agreement between the two graphs, and the same link (both ends of the link) in both graphs The greater the number of links having the same word in the node), the greater the degree of matching between the two graphs, and the greater the weight of the links in one of the graphs. , Calculate the degree of agreement between the graphs representing the subject so that the greater the weight of similar links in the other graph, the greater the degree of agreement between both graphs Including one of the calculation means.

According to the present invention (claim 9), the similarity determination means 4
At 50, each graph is divided into subgraphs based on how strongly the word sets used in the graph are related, and each subgraph is divided into subgraphs within the subgraph. If there is no link between any of the nodes, a link with a small weight is generated in the subgraph, the subgraph is recombined, the generated link is added to the subgraph as it is, and the graph before the division is returned. Second calculating means for calculating the degree of coincidence between the graphs expressing the subject.

According to the present invention (claim 10), the similarity determining means 450 compares each graph based on how strong the word sets used in the graph are related to each other. If the subgraph is divided into graphs and each subgraph has no link between any nodes between the subgraphs, a link having a small weight is generated in the subgraph, and a match is made for each subgraph. And a third calculating means for calculating the degree of.

According to the present invention (claim 11), the similarity determination means 450 compares each graph based on how strong the word sets used in the graph are related to each other. If the subgraph is divided into subgraphs and if there is no link between any nodes in the subgraph in each subgraph, a link having a small weight is generated in the subgraph, and the degree of matching for each subgraph is And calculating the sum of the degrees of matching calculated for each of the subgraphs, thereby calculating the degree of matching between the graphs expressing the subject.

FIG. 3 is a diagram illustrating the principle of the configuration of the document classification apparatus according to the present invention. The present invention (claim 12) provides a subject graph creating unit 610 that acquires a document corresponding to a document ID from a document storage unit 611 in which a document is stored, and creates a graph expressing the subject of the document. When a graph representing the subject of one document is input, a graph similarity determining unit 620 that determines the degree of coincidence between them, and a classifying unit 630 that classifies the document based on a matrix representing the similarity between the documents
And a classification control unit 640 for controlling the entire classification work.

According to the present invention (claim 13), in the subject graph creating means 610, in order to measure the degree of coincidence between graphs expressing the subject, a word string or a Boolean operator combination of word strings, a sentence or A word extraction unit that extracts words used in the document element from a document element formed of a document or a document set, and an importance assignment unit that assigns importance to each extracted word; Each document is represented by a relevance assigning means for assigning a relevance between any two extracted words and a graph in which the importance of the word is set as the weight of the node and the relevancy between the words is set as the weight of the link. Means for expressing the theme of the element, wherein the graph similarity determination means 620 determines the similarity between the document elements based on the degree of agreement between the graphs expressing the theme. No.

According to the present invention (claim 14), in the graph similarity determination means 620, the greater the number of similar nodes (nodes containing the same word) in both graphs, the greater the match between the graphs. If the degree is large and nodes in one graph are heavily weighted,
The greater the weight of similar nodes in the other graph, the greater the degree of agreement between the two graphs, and a similar link in both graphs (the word contained in the nodes at both ends of the link) The greater the number of links), the greater the degree of coincidence between the two graphs, and the greater the weight in the links in one graph, the greater the number of links in the other graph. The first calculating means for calculating the degree of coincidence between the graphs representing the subjects is such that the greater the weight of the similar link of the above, the greater the degree of coincidence between the two graphs. Including.

According to the present invention (claim 15), in the graph similarity determination means 620, each graph is determined based on how strong the word sets used in the graph are related to each other. , If each subgraph has no link between any nodes in the subgraph, generate a link with a smaller weight to the subgraph, rejoin the subgraphs, It includes a second calculating unit that adds the generated link to the graph as it is, returns the graph before the division, and calculates the degree of agreement between the graphs expressing the subject.

According to the present invention (claim 16), in the graph similarity determination means 620, each graph is determined based on how strong the word sets used in the graph are related to each other. If the subgraph is divided into subgraphs, and each subgraph has no link between any nodes between the subgraphs, a link having a small weight is generated in the subgraph, and a link is generated for each subgraph. A third calculating means for calculating the degree of coincidence is included.

According to the present invention (claim 17), in the graph similarity determination means 620, each graph is determined on the basis of how strong the word sets used in the graph are related to each other. If there is no link between any nodes in the subgraph in each subgraph, a link having a small weight is generated in the subgraph, and a match is determined for each subgraph. A fourth calculating means for calculating the degree of match between the graphs expressing the subject by calculating the degree and calculating the sum of the degrees of match calculated for each of the subgraphs.

The present invention (claim 18) is a storage medium storing a document search program for searching for a document based on a search request from a user. The storage medium analyzes a search request from a user and sets a search key. A search interface process for extracting, a search key subject graph creation process for generating a graph expressing the subject of the search key from the search key, and a word information management process for acquiring a set of document IDs of documents in which a specified word appears. , When the document ID is specified,
A search target document subject graph creation process for obtaining a document corresponding to the document ID from the search target document storage unit storing the search target document and creating a graph expressing the subject of the document; A similarity determination process of inputting the graph to be expressed and the graph expressing the subject of the search target document, and determining how similar they are;
It has a search interface process, a search key subject graph creation process, a word information management process, a search subject document subject graph creation process, and a search control process for controlling the similarity determination process.

According to the present invention (claim 19), in a search key subject graph creation process, a word string, a Boolean operator combination of a word string, or a document element composed of a sentence, a document, or a document set is used. A word extraction process for extracting words used in a document element, an importance assignment process for assigning importance to each extracted word, and a relevance between any two extracted words. The subject-of-search-document-graph creating process, the subject of each document element is represented by a graph in which the importance of a word is the weight of a node and the relevance between the words is the weight of a link. A similarity determination process including a theme representation process for representing the subject, wherein the similarity determination process determines a similarity between the document elements based on a degree of agreement between graphs representing the subject. Including.

According to the present invention (claim 20), in the similarity determination process, the greater the number of similar nodes (nodes containing the same word) in both graphs, the greater the degree of matching between the graphs. If the value is large and a node in one graph is heavily weighted, the greater the weight of a similar node in the other graph, the better the match between the two graphs The greater the number of similar links in both graphs (links having the same word in the nodes at both ends of the link), the greater the degree of matching between the two graphs. If a link in one graph is heavily weighted, the greater the weight of a similar link in the other graph, the greater the degree of matching between the two graphs. To include a first calculation process of calculating the degree of match between the graph representing the subject.

According to the present invention (claim 21), in the similarity determination process, each graph is partially divided based on how strongly the word sets used in the graph are related to each other. If the subgraph is divided into graphs and each subgraph has no link between any nodes in the subgraph, a link having a small weight is generated in the subgraph, the subgraph is rejoined, and the subgraph is recombined. A second calculation process of adding the generated link as it is and returning to the graph before division to calculate the degree of agreement between the graphs expressing the subject is included.

According to the present invention (claim 22), in the similarity determination process, each graph is divided into subgraphs based on how strong the word sets used in the graph are related to each other. If there is no link between any nodes between the subgraphs in each of the subgraphs, a link having a small weight is generated in the subgraph, and a match is determined for each subgraph. And a third calculation process for calculating the degree.

According to the present invention (claim 23), in the process of determining the similarity between document elements, each graph is determined based on how strong the word sets used in the graph are related to each other. Then, if there is no link between any nodes in the subgraph in each subgraph, a link having a small weight is generated in the subgraph, and the subgraph is matched for each subgraph. And calculating the sum of the degrees of matching calculated for each of the subgraphs, thereby calculating the degree of matching between the graphs representing the subjects.

The present invention (claim 24) is a storage medium storing a document classification program mounted on a document classification device for classifying documents, wherein the document storage means in which the document is stored is provided with a document ID. The subject graph creation process of acquiring a document corresponding to the subject and creating a graph representing the subject of the document, and a graph similarity determination process for determining the degree of coincidence between the two when a graph representing the subject of two documents is input Based on the degree determination process and a matrix representing the similarity between documents,
A classification process for classifying the document; and a classification control process for controlling the entire classification work of the classification process.

According to the present invention (claim 25), in a subject graph creation process, a word string or a Boolean operator combination of a word string or a sentence or a character string is used to measure the degree of coincidence between graphs representing the subject. A word extraction process for extracting words used in a document element from a document or a document element composed of a document set, an importance assignment process for assigning importance to each extracted word, The relevance assigning process of assigning a relevance between any two given words and a graph in which the importance of the word is set as the weight of the node and the relevancy between the words is set as the weight of the link are used for each document element. And a document element similarity determination process for determining the similarity between document elements based on the degree of agreement between the graphs representing the themes. .

According to the present invention (claim 26), in the graph similarity determination process, as the number of similar nodes (nodes containing the same word) in both graphs increases,
The degree of coincidence between the graphs is set to a large value, and if a node in one graph has a large weight, the larger the similar node in the other graph has a large weight, the more both The degree of coincidence between the graphs is set to a large value, and the greater the number of similar links (links having the same word in the nodes at both ends of the link) in both graphs, the greater the degree of coincidence between the two graphs If the link in one graph is heavily weighted, the greater the weight of a similar link in the other graph, the greater the match between the two graphs A first calculation process of calculating the degree of agreement between the graphs representing the subjects so as to increase the degree of.

According to the present invention (claim 27), in the graph similarity determination process, each graph is determined based on how strongly the word sets used in the graph are related to each other. If the subgraph is divided into subgraphs and each subgraph has no link between any nodes in the subgraph, a link having a small weight is generated in the subgraph, and the subgraph is recombined. The second is to add the generated link as it is and return to the graph before division to calculate the degree of agreement between the graphs expressing the subject.
Including the calculation process.

According to the present invention (claim 28), in the graph similarity determination process, each graph is divided into parts based on how strongly the word sets used in the graph are related to each other. If the subgraph is divided into graphs and each subgraph has no link between any nodes between the subgraphs, a link having a small weight is generated in the subgraph, and a match is made for each subgraph. A third calculation process for calculating the degree of

According to the present invention (claim 29), in the graph similarity determination process, each graph is determined based on how strongly the word sets used in the graph are related to each other. If the subgraph is divided into subgraphs and if there is no link between any nodes in the subgraph in each subgraph, a link having a small weight is generated in the subgraph, and the degree of matching for each subgraph is And calculating the sum of the degrees of matching calculated for each subgraph, thereby calculating the degree of matching between the graphs representing the subjects.

As described above, according to the present invention, the problem that the similarity between document elements having a plurality of themes and subtitles cannot be determined with high accuracy is determined by using the relevance between words. It is possible to determine the similarity between document elements with high accuracy. For example, in the example of the problem to be solved by the above-described invention, if the user wants to search for a document related to "robot performing information search", in the present invention, "information search" and "robot" are strongly related. A document has a higher similarity than a document that does not. In a document having two themes, "information retrieval system" and "industrial robot", "information retrieval" and "robot" are not strongly related to each other. No. As described above, according to the present invention, it is possible to accurately determine the similarity.

Further, according to the present invention, the problem that the similarity cannot be determined using the characteristics of the document element cannot be determined between words having a strong dependency relationship in a sentence or words included in the same paragraph. Since a high degree of relevance can be given between them, the problem can be solved by determining the similarity using these features. As described above, the similarity can be accurately determined by using the present invention.

Further, to solve the problem that the accuracy of the determination of the similarity is reduced due to the failure of the word extraction due to the use of the morphological analysis, in the present invention, the above-mentioned failure of the extraction of the word "Internet" is considered. Using an example,
Since morphological analysis is used, similar to the above example,
The word “internet” is not extracted, and this word is extracted as two words “inter” and “net”. However, when a certain document element has a character string “Internet”, there is a strong relationship between the extracted words “inter” and “net” in the document element. Therefore, the document element in which the character string “Internet” appears has a higher similarity than the document element in which “inter” and “net” appear separately. Therefore, even if the morphological analysis fails, the present invention can prevent a decrease in the accuracy of the similarity determination.

As described above, according to the present invention, since the similarity between document elements can be determined with high accuracy, a highly accurate similarity determination method, a document search device, a document classification device, a storage medium storing a document search program, and a document classification program A storage medium storing a program can be provided.

[0048]

FIG. 4 is a flow chart for explaining a similarity determination method according to the present invention. Step 10) Extract words used in the document element. Step 20) Calculate the importance of each word used in the document element.

Step 30) The degree of relevance between any two words used in the document element is calculated. Step 40) The subject of each document element is expressed by a graph in which the importance of a word is set as the weight of a node and the degree of association between words is set as the weight of a link. Hereinafter, this graph is called a theme graph.

Step 50) The similarity between the document elements is determined based on the degree of coincidence between the subject graphs generated in this way. Hereinafter, the operation of each of the above steps will be described in detail. Step 10) Word extraction: Word extraction is performed by morphological analysis of document elements. An existing technology is used for the morphological analysis method.

Step 20) Calculation of importance: The importance of words used in each document element is calculated as follows. In the present invention, as a document element, a word string, a Boolean operator combination of words, a sentence, a document, and a document set are assumed, and a method of calculating the importance of a word for each is described below.

Word sequence: The importance is determined by setting the importance of all words to the same value or by letting the user explicitly specify the importance of each word. Boolean operator combination of words: The importance is determined in the same manner as in the case of word strings. Sentence: All words have the same importance, or the importance is determined according to the part of speech of a word (e.g., a proper noun is given a higher importance than an adverb).

Document: The importance is determined in the same manner as in the case of a sentence, or word position information (words appearing in a title are given high importance), appearance frequency information (high occurrences) The calculation is performed based on a word having a high frequency, a high degree of importance, and the number of document elements appearing in the entire document element set (a high degree of importance is given to a word that appears only in a specific document element).

Document set: Considering the entire document set as one large document (a document in which all documents are combined), the importance is calculated in the same manner as in the case of a document. Step 30) Calculation of relevance: The relevancy between words used in each document element is calculated as follows. In the present invention, as a document element, a word string, a combination of Boolean operators of words, a sentence, a document, and a set of documents are assumed.

Word string: all 2 included in the document element
The relevance is determined by making the relevance between words equal or by having the user explicitly specify the relevance. -Word Boolean operator combination: In addition to the word string method, the degree of relevance is determined according to the type of Boolean operator. For example, words connected by and have a greater degree of relevance than words connected by or.

Sentence: A sentence is calculated by using a method using a word string or by using the following dependency information.
First, the dependency relation of the sentence is analyzed. An existing technique is used as a method of analyzing the dependency relationship. It is assumed that those having a direct dependency relationship have a strong relationship, and those having an indirect dependency relationship have a weak relationship. For example, if there is a sentence “Use word importance for information search”, the relevance of “information” and “search” is larger than the relevance of “information” and “word”. I do. This is because “information” and “search” have a direct dependency relationship, whereas “information” and “word” do not have a direct dependency relationship.

Document: The degree of relevance between words is calculated by using a sentence method or by one of the following two methods. -Use of co-occurrence information: If two words co-occur in the same sentence (or within the specified number of characters), count the number of co-occurrences. The more co-occurrences, the more
The degree of association between these two words is set to a large value.

Using structure information: Analyze the structure (chapter, paragraph, etc.) of the document. Since words appearing in a paragraph are related to the headword of the paragraph, and words in the paragraph are considered related,
The degree of relevance between words is determined based on the frequency information in the paragraph. For example, words that occur frequently only in a paragraph are strongly related to the headword of that section,
It is also assumed that those words have a strong relationship.

Document set: A document set is considered as one large document (a document in which all documents are combined), and the degree of relevance is calculated in the same manner as in the case of a document. Step 40) Creation of a subject graph: A graph is created in which the importance of the word obtained in step 20 is used as a node weight, and the degree of association between the words obtained in step 30 is used as a link weight. This graph (subject graph) expresses the subject of the document element.

Step 50) The similarity between the document elements is determined by measuring the degree of coincidence between the subject graphs of the document elements created in step 140. FIG. 5 shows the configuration of the similarity determination processing. FIG. 5 is a diagram for explaining the operation of the similarity determination processing of the present invention. 1. Graph coincidence measurement processing (steps 111 and 12)
3, 133, 143): The degree of coincidence between the subject graphs q and u of the document element is calculated by the following equation. The importance of the words used in the graphs q and u is represented by the following vectors, respectively.

V _q = (v _q1 , v _q2 ,..., V _qn ) (1) v _u = (v _u1 , v _u2 ,..., V _un ) (2) where v _qi and v _ui are respectively It represents the importance of the word i in the document element q and the importance of the word i in the document u. Dot product f _{v of} these vectors

[0062]

(Equation 1)

Is calculated. The relevance between words used in the graph q and the graph u is represented as follows. _{r q = (v q11, v} q12, ..., v q21, v q22, ..., v qnn) (3) r u = (v u11, v u12, ..., v u21, v u22, ..., v unn) ( 4) Here, v _qij and v _uij represent the relevance between the word i and the word j in the document element q and the relevance between the word i and the word j in the document element u, respectively.

The inner product of these vectors

[0065]

(Equation 2)

Is calculated. obtained from f _v and f _r as follows coincidence degree between the graph. Matching degree = f _v ^p * f _r ^q where p and q are positive constants. 2. Graph division processing (steps 121, 131, and 14)
1): The graph is divided by the following processing, and a link having a small weight is generated in each subgraph.

( _A ) The graph G _A is divided into p subgraphs by G _Ai (i = 0, 1,
.., P). Here, the strength of the coupling force is, for example, "Any node in each subgraph has
Make sure there is a link with distance 1 or distance n
There are m (m ≧ 1) or more links (n ≧ 2) or less. And so on. Here, the distance between the node a and the node b is the number of links that pass from the node a to the node b.

(B) If no link exists between arbitrary nodes in the divided subgraph, a link having a weak weight is generated between these nodes. FIG. 6 shows a graph division process when n = 2 and m = 2. In this example, the graph G _A (210) before the division is composed of three subgraphs (G
_A1 (221), G _A2 (222), and G _A3 (233)). The division is as follows: For G _A1 (211), nodes AB, AC, BD,
A link with a distance of 1 exists between CDs, and two links with a distance of 2 exist between each of AD and BC. For G _A2 (222), a link with a distance of 1 exists between nodes BD, BE, and DE. For G _A3 (223), there is a link with a distance of 1 between nodes DF.

For this reason, the condition that "a link with a distance of 1 always exists or two or more links with a distance of 2 or less exist between any two nodes in each subgraph" is satisfied. Because. Also, G _A1 (2
A broken line in 21) is a link with a weak weight added in (b) of the graph division processing. As is apparent from this processing, the words in the divided subgraphs are connected with a strong connection force. Therefore, these subgraphs are
Since the document is composed of a set of words that are semantically related, the sub-documents composed of these subgraphs each represent a subtitle of the original document.

By generating links in the subgraph in this way, it is expressed on the graph that the words included in the respective subtitles have a certain degree of association. 3. Graph reassociation processing (step 122): the partial graph G _Ai (i = 0, 1,..., P) created by the graph division processing
Is recombined with the graph G ′ _A before division. At this time, G '
_A is obtained by adding a link generated in the graph division processing (step 121).

FIG. 7 shows an example of the graph recombining process. In this example, three subgraphs (G _A1 (311), G _A2 (312), G _A3 ) created from G _A (210) in FIG.
(313)) is recombined into the original graph to create G ′ _A (320). To G ' _A created by the graph division processing and the graph recombining processing, a link with a weak weight that did not exist in G _A between AD and BC is added.

4. Subgraph coincidence measurement processing (steps 132 and 142): The degree of coincidence is measured using the intergraph coincidence measurement processing for each of the partial graphs created by the graph division processing. 5. Matching degree total processing (step 144): The value obtained by summing up the matching degrees obtained for each subgraph is set as the matching degree of the entire graph before division.

Next, the method of calculating the similarity shown in FIG.
) Will be described. 1. Method not using graph division (step 110): (a) Pass the subject graphs G _q and _Gu of the document element to the inter-graph coincidence measurement processing (step 111).

(B) In the inter-graph coincidence measuring process (step 111), the coincidence between these two graphs G _q and _Gu is measured and output. The degree of coincidence between subject graphs obtained by this method is defined as the degree of similarity between document elements. This method is fast because no graph division is used. 2. Method using graph division and recombination (step 12
0): (a) Pass the subject graphs _Gq and _Gu of the document element to the graph division processing (step 121).

(B) Graph division processing (step 12)
1) sets G _q and _Gu to a plurality of subgraphs G
_qi (i = 0, 1, ... p), _Guj (j = 0, 1, ..., r)
If there is no link between any of the nodes in G _qi and G _uj , a link having a small weight is generated between these nodes. (C) The graph recombining process (step 122) includes the subgraph G created in the graph dividing process (step 121).
G ′ _q and G ′ _u are created by recombining _qi and G _uj to the state before the graph division. As described above, at the time of reconnection, the links generated in the respective subgraphs are added to G ′ _q and G ′ _u .

[0076 (d) The graph between coincidence measurement process (step 123), measures and outputs the coincidence degree between G _'q and G' _u. The degree of matching between the subject graphs obtained in this way is defined as the degree of similarity between document elements. In this method, similarity can be determined using indirect word-to-word associations (words included in the same subtitle have some direct association), so that a more accurate similarity can be determined. Can be calculated.

3. Method for measuring the degree of coincidence for each subgraph (step 130): (a) Pass the subject graphs _Gq and _Gu of the document element to a graph division process (step 131). (B) The graph division processing (step 131)
G ′ _q and G ′ _u are _converted into a plurality of subgraphs G _qi (i = 0, 1,
, P) and G _uj (j = 0, 1,..., R)
If there is no link between any nodes in G _qi and G _uj , a link having a small weight is added between these nodes.

(C) The subgraph coincidence measuring process (step 132) measures the coincidence of all the combinations of the subgraphs _Gqi and _Guj . At this time, the inter-graph coincidence measurement processing (step 133) is used.
Outputs the degree of coincidence obtained for each subgraph. The degree of coincidence for each subgraph obtained in this way is defined as the degree of similarity for each divided sub-document (subtitle). In this method, the similarity can be calculated for each sub-document (subtitle) in the document, not for the entire document.

4. Method using the sum of the degrees of coincidence for each subgraph (step 140): (a) Each of the subject graphs G _q and _Gu of the document element is passed to the graph division processing (step 141). (B) The graph division processing (step 141) is performed by G _q ,
The G _u plurality of subgraphs _{G qi (i = 0,1, ...} , p),
G _uj (j = 0, 1,..., R), and if there is no link between any of the nodes in G _qi and G _uj , a link having a small weight is added between these nodes.

(C) The subgraph coincidence measuring process (step 142) measures the coincidence of all combinations of the subgraphs _Gqi and _Guj . On this occasion,
The graph matching degree measurement processing (step 143) is used. (D) The matching degree total processing (step 144) calculates and outputs the sum of the matching degrees for all the partial graphs obtained in the subgraph matching degree measuring processing (step 142).

The sum of the coincidences determined in this way is regarded as the similarity between the document elements. According to this method, the similarity between document elements can be determined based on the sum of the similarities for each subtitle in the document, so that more accurate calculation of the similarity can be performed. To summarize the above, the method that can perform the processing at high speed is the “method without using graph division (step 110)”.

In the "method using graph division and recombination (step 120)", similarity determination using indirect word association is performed instead of performing complicated processing called graph division processing. it can. In the "method of measuring the degree of coincidence for each subgraph (step 130)", the document element is not treated as one document, but the similarity is determined for each independent sub-document divided for each subtitle. It can be performed.

In the "method using the sum of the degrees of coincidence for each subgraph (step 140)", the processing is extremely complicated, but the similarity of the entire document is obtained as the sum of the degrees of similarity for each subtitle in the document. Therefore, more accurate similarity can be determined.

[0084]

Embodiments of the present invention will be described below with reference to the drawings. First, a document search device using the above-described method will be described. FIG. 8 shows the configuration of a document search device according to one embodiment of the present invention. The document search device shown in the figure includes a search interface unit 410, a search key subject graph creation unit 420, a word information management unit 430, and a search subject subject graph creation unit 44.
0, similarity determination unit 450, search control unit 460, search target document database 441, and index file 4
31.

The search interface unit 410 analyzes a search request from a user, extracts a search key, and passes it to the search control unit 460. Further, the search result is received from the search control unit 460 and returned to the user. The search key subject graph creation unit 420 creates a subject graph from the search key. The word information management unit 430 refers to the index file 431 to obtain the document ID in which the specified word appears.
Get the set of Here, the index file 43
A hash table 1 has a word as a key and a set of document IDs in which the word appears as a value.

The search subject document subject graph creation unit 440
When a document ID is specified, a document corresponding to the document ID is acquired from the search target document database 441, and a subject graph of the document is created. The similarity determination unit 450 receives the subject graph of the search key and the subject graph of the search target document, and determines the similarity between them. Here, the method of determining the similarity uses the method of the flowchart of FIG. 5 described above.

The search control unit 460 performs the following processing. (A) Obtain a search key from the search interface unit 410. (B) From the search key subject graph creation unit 420, a subject graph created from this search key is obtained. (C) From the word information management unit 430, a set of document IDs in which any one of the words in the subject graph appears is acquired.

(D) The following processing is executed for each element of the set of document IDs. The subject graph of the search target document corresponding to the document ID is acquired from the search subject document subject graph creation unit 440. The similarity between the subject graph of the search target document and the subject graph of the search key is acquired from the similarity determination 450.

(E) The set of document IDs is sorted in descending order of similarity, and documents corresponding to the top n document IDs are set as search results. Hereinafter, the flow of processing will be described using the following example. FIG. 9 shows an example of creating a subject graph according to one embodiment of the present invention. Search key Q: (information or document) and search Search target document U: A document consisting of the following five sentences “About the subject of information.

About the subject of the document. Improve search efficiency. Search for information. Search for documents. Step 501) The user inputs a search request.

Step 502) The search interface unit 410 extracts a search key from the search request input by the user, and passes it to the search control unit 460. Step 503) The search control unit 460 passes the search key to the search key subject graph creation unit 420. Step 504) Search key subject graph creation section 420
Creates a subject graph from a search key,
Pass to 0. In this example, the subject graph 510 of the search key shown in FIG. However, the importance of all words was set to 1.0, and the relevance between words was set to 0.5 for "or" and 1.0 for "and".

Step 505) The search control section 460
Each word used in the subject graph of the search key is passed to the word information management unit 430. Step 506) The word information management unit 430 acquires from the index file 431 a set of document IDs in which the word appears at least once, and passes it to the search control unit 460.

Step 507) The search control unit 460
Each element of the set of document IDs acquired from the word information management unit 430 is passed to the search subject document graph creation unit 440. Step 508) Search subject document subject graph creation unit 44
0 acquires the document corresponding to the document ID from the search target document database 441, creates a subject graph of the document, and passes it to the search control unit 460. In this example, the subject graph 520 of the search target document in FIG. However, the importance of the word was a value proportional to the number of appearances, the relevance between words was a value proportional to the number of co-occurrence in the sentence, and unnecessary words were removed.

Step 509) The search control unit 460
The subject graphs obtained from the search subject document subject graph creation unit 440 and the subject graph of the search key are passed to the similarity determination unit 450. Step 510) The similarity determination unit 450 determines the similarity between the subject graph of the search key and the subject graph of each search target document.

This time, a method of determining the similarity will be described as an example of the case where the “method without using graph division” is used.
First, a vector representing the importance of the following words is generated from the subject graph 510 of the search key and the subject graph 520 of the search target document in FIG. However, the elements of the vectors v _Q and v _U indicate the importance of (information, document, search, subject, efficiency) in the search key Q and the search target document U, respectively, and the importance of words not present in the graph. Was set to 0.0.

V _Q = (1.0, 1.0, 1.0, 0.0, 0.0) (5) v _u = (0.6, 0.6, 1.0, 0.6, 0.3) (6) The inner product f _v of these vectors is f _v = 2 .2.

Next, similarly, a vector representing the degree of association is generated. Where the elements of the vectors r _Q and r _U are
The search key Q, (information and subject, information and document, information and search, information and efficiency, subject and document, subject and search, subject and efficiency, document and search, document and efficiency, search and Efficiency), and the weight of a link that does not exist in the graph is set to 0.

R _Q = (0.0, 0.5, 1.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0) (7) r _U = (1.0, 0.0, 1.0, 0.0, 1.0, 0.0, 0.0, 1.0, 0.0, 1.0) (8) The inner product f _v of these vectors is f _v = 2.0. Therefore, if p = 1 and q = 1, the subject graph 510 of the search key and the subject graph 520 of the search target document
The degree of coincidence, coincidence degree _{= f U * f r = 2.2} * 2.0 = 4.4 , and the similarity of the search key Q searched documents U in this example, 4.4 and calculations Is done.

Step 511) The search control unit 460
The document set is rearranged in descending order based on the similarity of each document acquired from the similarity determination unit 450, and the top n items are set as search results, which are passed to the search interface unit 410. Here, when the similarity determination unit 450 uses the method of measuring the degree of coincidence for each subgraph, the top n sub-documents are the search results.

Step 512) The search interface unit 410 returns a search result to the user. In this example,
(Document U, 4.4) is the search result. Next, a document classification apparatus using the present invention will be described.
FIG. 10 shows the configuration of a document classification device according to an embodiment of the present invention. The document classification device shown in FIG.
0, similarity determination section 620, classification section 630, classification control section 6
40, a document database 611.

In the figure, the subject graph creation unit 610
Is exactly the same as the search subject document graph creation unit 440 of the document search device described above. Thematic graph creation unit 61
0 acquires the document corresponding to the document ID from the document database 611 and creates a subject graph of the document. When the subject graphs of two documents are input, the similarity determination unit 620 determines the similarity between the two. Here, the similarity determination is performed using the similarity determination method similar to the similarity determination unit 450 of the above-described document search device.

The classifying section 630 classifies documents based on the inter-document similarity matrix. Here, the inter-document similarity matrix has the following format. Document 1 Document 2… Document n Document 1 s ₁₁ s ₁₂ … s _1n Document 2 s ₂₁ s ₂₂ … s _2n ………… Document ns s _n1 s _n2 … s _nn where s _ij is document i and document j S _ij =
s _ji and s _ii are infinite.

A classification method when an inter-document similarity matrix is given is, for example, clustering in which documents having the highest similarity are sequentially combined. The specific classification method is based on the existing description. The classification control unit 640 controls the entire classification work. A series of operations of the above configuration will be described below.

Step 601) The user designates how many document sets the documents in the document database 611 are to be classified into (classification number). Step 602) The classification control unit 640 passes the document IDs of all the documents included in the document database 611 to the subject graph creation unit 610. Step 603) The subject graph creation unit 610 stores the document corresponding to each document ID in the document database 611.
, A theme graph is created, and the subject graph is created.
Pass to 40.

Step 604) The classification control unit 640
All two combinations of the theme graphs obtained from the theme graph creation unit 610 are passed to the similarity determination unit 620. Step 605) The similarity determination unit 620 determines the similarity between the subject graphs, and passes the similarity to the classification control unit 640. Step 606) The classification control unit 640 creates an inter-document similarity matrix from the similarities between the documents of all the two combinations acquired from the similarity determination unit 620, and sends it to the classification unit 630 together with the number of classifications input by the user. hand over.

Step 607) The classification unit 630 classifies the document set based on the inter-document similarity matrix and the number of classifications.
The classification result is passed to the classification control unit 640. Step 608) The classification control unit 640 returns a classification result to the user. Further, the components of the document search device and the document classification device in the above embodiment are constructed as a program,
It is easily stored in a disk device connected to a computer used as a document search device and a document classification device, or in a portable storage medium such as a floppy disk or a CD-ROM, and installed when implementing the present invention. Thus, the present invention can be realized.

The present invention is not limited to the above-described embodiments, but can be variously modified and applied within the scope of the claims.

[0108]

As described above, according to the present invention, the similarity between document elements can be determined based not only on the importance of words but also on the degree of relevance between words. The determination of the degree of similarity can be performed. In addition, the similarity between the search key and the search target document is calculated using not only the search key and the importance of the word in the search target document but also the relevance between the search key and the word in the search target document. Therefore, even if the search key is a sentence or a document, or if the search target is the whole document, a more accurate information search can be expressed.

Similarly, the similarity between documents can be calculated using not only the importance of the words in the document but also the relevance between the words in the document, so that a more accurate document classification can be realized. it can.

[Brief description of the drawings]

FIG. 1 is a diagram for explaining the principle of the present invention.

FIG. 2 is a diagram illustrating the principle configuration of a document search device according to the present invention.

FIG. 3 is a diagram illustrating the principle configuration of a document classification device according to the present invention.

FIG. 4 is a flowchart illustrating a similarity determination method according to the present invention.

FIG. 5 is a diagram illustrating an operation of a similarity determination process according to the present invention.

FIG. 6 is a diagram illustrating a graph classification process according to the present invention.

FIG. 7 is a diagram for explaining a graph recombining process according to the present invention.

FIG. 8 is a configuration diagram of a document search device according to an embodiment of the present invention.

FIG. 9 is a diagram showing an example of creating a subject graph according to one embodiment of the present invention.

FIG. 10 is a configuration diagram of a document classification device according to an embodiment of the present invention.

[Explanation of symbols]

210 Graph before division 211, 222, 223 Subgraph 311, 312, 313 Subgraph 320 Rejoined graph 410 Search interface unit, search interface unit 420 Search key subject graph creation unit, search key subject graph creation unit 430 Word information management Means, word information management unit 431 Index file 440 Search target document subject graph creation unit, Search target document subject graph creation unit 441 Search target document storage unit, Search target document database 450 Similarity determination unit, Similarity determination unit 460 Search control unit Search control unit 510 subject graph of search key Q 520 subject graph of search target document U 610 subject graph creation unit, subject graph creation unit 611 document storage unit, document database 620 graph similarity determination unit, similarity determination unit 630 Classification unit, classification unit 640 Classification control unit, classification unit control unit

Claims (29)

[Claims] 1. A similarity determination method for appropriately determining similarity between document elements, comprising: a word string, a Boolean operator combination of word strings, a sentence, a document, or a document set Extracting words used in the document element from the elements; assigning importance to each of the extracted words; assigning a relevance between any two extracted words; The subject of each document element is expressed by a graph in which importance is set as the weight of the node and the relevance between the words is set as the weight of the link. Based on the degree of matching between the graphs expressing the subject, the document element is used. A similarity determination method characterized by determining a similarity between the two. 2. When calculating the degree of coincidence between graphs for expressing the subject, the greater the number of similar nodes (nodes containing the same word) in both graphs, the more the graphs If the degree of matching between them is large, and if a node in one graph is heavily weighted, the greater the weight of a similar node in the other graph, the greater the weight of both graphs The degree of matching between the two graphs is set to a large value, and the greater the number of similar links (links having the same word included in the nodes at both ends of the link) in both graphs, the greater the number of matches between the two graphs If the degree is set to a large value and a link in one graph is heavily weighted, the greater the weight of a similar link in the other graph, the greater the match between the two graphs Degree The to a large value, the similarity judging method of claim 1 wherein calculating a degree of match between the graph representing the subject. 3. When calculating the degree of coincidence between graphs for expressing the subject, each graph is associated with a word set used in the graph with a certain degree of strength. If there is no link between any nodes in the subgraph in each of the subgraphs, a link with a small weight is generated in the subgraph, and the subgraph is divided into subgraphs. Is recombined, and the link generated in the subgraph is added as it is,
2. The similarity determination method according to claim 1, wherein the degree of coincidence between the graphs expressing the subject is calculated by returning to the graph before the division. 4. When calculating the degree of coincidence between graphs for expressing the subject, how each word is used in the graphs is related to each graph. On the basis of the,
If the subgraph is divided into subgraphs, and if each of the subgraphs has no link between any nodes between the subgraphs, a link having a small weight is generated in the subgraph; 2. The similarity determination method according to claim 1, wherein the degree of coincidence is calculated for each case. 5. When calculating the degree of coincidence between graphs for expressing the subject, each graph is associated with a word set used in the graph with a certain degree of strength. If there is no link between any nodes in the subgraph in each of the subgraphs, a link having a small weight is generated in the subgraph, and each subgraph is divided into subgraphs. The similarity according to claim 1, wherein a degree of coincidence is calculated for each graph, and a degree of coincidence between graphs expressing the subject is calculated by calculating a sum of degrees of coincidence calculated for each of the subgraphs. Judgment method. 6. A document search device for searching a document based on a search request from a user, comprising: a search interface unit for analyzing a search request from the user and extracting a search key; Search key subject graph creating means for generating a graph expressing the subject of the search key; word information management means for acquiring a set of document IDs of documents in which the specified word appears; and when the document ID is specified, A search target document subject graph creating unit for obtaining a document corresponding to the document ID from the search target document storage unit and creating a graph expressing the subject of the document; a graph expressing the subject of the search key and a search target document A similarity determination unit for inputting a graph expressing the subject of the search, and determining how similar they are; the search interface unit; A document search device, comprising: a title graph creation unit, the word information management unit, the search target document subject graph creation unit, and a search control unit that controls the similarity determination unit. 7. The search key subject graph creating means is used in the document element from a word string, a Boolean operator combination of the word strings, or a document element composed of a sentence, a document, or a document set. Word extracting means for extracting a word which is present, importance assigning means for assigning importance to each of the extracted words, and relevance assigning means for assigning a relevance between any two extracted words. The subject expression representing the subject of each document element by a graph in which the importance of the word is used as the weight of the node and the relevance between the words is used as the weight of the link. 7. The document search apparatus according to claim 6, further comprising: means for determining a similarity between the document elements based on a degree of coincidence between graphs expressing the subject. 8. The similarity determination means sets the degree of coincidence between the graphs to a larger value as the number of similar nodes (nodes including the same word) in both graphs increases. If a node in the graph is heavily weighted, the greater the weight of a similar node in the other graph, the greater the degree of matching between both graphs; The greater the number of similar links (links having the same word in the nodes at both ends of the link) of the graph, the greater the degree of matching between the two graphs is If a link is heavily weighted, the subject may be modified such that the greater the weight of a similar link in the other graph, the greater the degree of agreement between the two graphs will be. Express Claim including a first calculating means for calculating a degree of match between the graph 6
Document search device as described. 9. The similarity determination unit divides each graph into subgraphs based on how strong the word sets used in the graph are related to each other. If there is no link between any nodes in the subgraph in the subgraph, a link having a small weight is generated in the subgraph, the subgraph is recombined, and the link generated in the subgraph is generated. As it is,
7. The document search apparatus according to claim 6, further comprising a second calculating unit that returns the graph before the division and calculates the degree of coincidence between the graphs expressing the subject. 10. The similarity determination means, based on how strong the word sets used in the graph are related to each other,
If the subgraph is divided into subgraphs, and if each of the subgraphs has no link between any nodes between the subgraphs, a link having a small weight is generated in the subgraph; 7. The document search apparatus according to claim 6, further comprising third calculating means for calculating the degree of matching for each of the documents. 11. The similarity determination means divides each graph into subgraphs based on how strong the word sets used in the graph are related to each other. If the subgraph has no link between any nodes in the subgraph, a link having a small weight is generated in the subgraph, and the degree of matching is calculated for each of the subgraphs. 7. The document search apparatus according to claim 6, further comprising: fourth calculating means for calculating a degree of coincidence between graphs expressing the subject by calculating a sum of degrees of coincidence calculated for each subgraph. 12. A subject graph creating means for acquiring a document corresponding to a document ID from a document storage means in which a document is stored, and creating a graph representing the subject of the document, and expressing the subject of the two documents. When a graph to be input is input, a graph similarity determination unit that determines the degree of coincidence between them, a classification unit that classifies the documents based on a matrix that represents the similarity between the documents, And a classification control unit for performing the classification. 13. The subject graph creating means, for measuring the degree of coincidence between graphs expressing the subject, by word strings, Boolean operator combinations of word strings, sentences, documents, or document sets. Word extracting means for extracting words used in the document element from the composed document element; importance assigning means for assigning importance to each of the extracted words; A relevance assigning unit for assigning a relevance between two words, and expressing the subject of each document element by a graph in which the importance of the word is used as a node weight and the relevance between the words is used as a link weight. 13. The theme similarity determining means, wherein the graph similarity determining means includes means for determining similarity between the document elements based on a degree of coincidence between graphs expressing the theme.
Document classification device as described. 14. The graph similarity determination means sets the degree of coincidence between the graphs to a larger value as the number of similar nodes (nodes including the same word) in both graphs increases. If a node in the graph has a large weight, the greater the weight of a similar node in the other graph, the greater the degree of matching between the two graphs, The greater the number of similar links (links containing the same word in the nodes at both ends of the link) in both graphs, the greater the degree of matching between the two graphs, If a certain link has a large weight, the more the similar link in the other graph has a larger weight, the greater the degree of matching between the two graphs will be. Claim 1 including a first calculating means for calculating a degree of match between the graph representing
2. The document classification device according to 2. 15. The graph similarity determination means divides each graph into subgraphs based on how strong the word sets used in the graph are related to each other, When there is no link between any nodes in the subgraph in each of the subgraphs, a link having a small weight is generated in the subgraph, the subgraph is recombined, and the subgraph is generated in the subgraph. Add the link as is,
13. The image processing apparatus according to claim 12, further comprising a second calculating unit configured to calculate a degree of coincidence between the graphs expressing the subject by returning to the graph before the division.
Document classification device as described. 16. The graph similarity determination means may determine each graph based on how strong the word sets used in the graph are related to each other.
If the subgraph is divided into subgraphs, and if each of the subgraphs has no link between any nodes between the subgraphs, a link having a small weight is generated in the subgraph; 13. The document classification device according to claim 12, further comprising a third calculating unit that calculates a degree of coincidence for each document. 17. The graph similarity determination means divides each graph into subgraphs based on how strong the word sets used in the graph are related to each other, If each of the subgraphs has no link between any nodes in the subgraph, generate a link with a small weight in the subgraph, calculate the degree of matching for each of the subgraphs, 13. The document classification device according to claim 12, further comprising: fourth calculating means for calculating a degree of coincidence between graphs expressing the subject by calculating a sum of degrees of coincidence calculated for each of the subgraphs. 18. A storage medium storing a document search program for searching for a document based on a search request from a user, comprising: a search interface process for analyzing a search request from the user and extracting a search key; A search key subject graph creation process for generating a graph representing the subject of the search key from the search key; a word information management process for acquiring a set of document IDs of documents in which a specified word appears; When specified, a search target document subject graph creation process for obtaining a document corresponding to the document ID from the search target document storage means storing the search target document and creating a graph expressing the subject of the document; Input a graph expressing the subject of the search key and a graph expressing the subject of the search target document, and determine how similar they are A similarity determination process for determining; a search interface process; a search key subject graph creation process; a word information management process; a search target document subject graph creation process; and a search control process for controlling the similarity determination process. A storage medium storing a document search program characterized by having: 19. The search key subject graph creation process is executed in a document element from a word string, a Boolean operator combination of word strings, or a document element composed of a sentence, a document, or a document set. A word extraction process for extracting a word, an importance assignment process for assigning importance to each extracted word, and a relevance assignment process for assigning a relevance between any two extracted words. The subject expression representing the subject of each document element by a graph in which the importance of the word is used as the weight of the node and the relevance between the words is used as the weight of the link. A similarity determination process, wherein the similarity determination process determines a similarity between the document elements based on a degree of agreement between the graphs representing the themes. Storage medium storing a document retrieval program according to claim 18, comprising a. 20. The similarity determination process, wherein the greater the number of similar nodes (nodes containing the same word) in both graphs, the greater the degree of matching between the graphs, If a node in the graph is heavily weighted, the greater the weight of a similar node in the other graph, the greater the degree of matching between both graphs; The greater the number of similar links (links having the same word in the nodes at both ends of the link) of the graph, the greater the degree of matching between the two graphs is If a link is heavily weighted, the subject may be modified such that the greater the weight of a similar link in the other graph, the greater the degree of agreement between the two graphs will be. First storage medium storing a document retrieval program according to claim 18, further comprising a calculation process of calculating the degree of match between the graph of current. 21. The similarity determination process divides each graph into subgraphs based on how strongly the word sets used in the graph are related to each other. If there is no link between any nodes in the subgraph in the subgraph, a link having a small weight is generated in the subgraph, the subgraph is recombined, and the link generated in the subgraph is generated. As it is,
19. The storage medium storing the document search program according to claim 18, further comprising a second calculation process for calculating the degree of coincidence between the graphs expressing the subject by returning to the graph before the division. 22. The similarity determination process may include rewriting each graph based on how strongly the word sets used in the graph are related to each other.
If the subgraph is divided into subgraphs, and if each of the subgraphs has no link between any nodes between the subgraphs, a link having a small weight is generated in the subgraph; 19. The storage medium storing the document search program according to claim 18, including a third calculation process for calculating a degree of coincidence every time. 23. The similarity determination process divides each graph into subgraphs based on how strong the word sets used in the graph are related to each other. If the subgraph has no link between any nodes in the subgraph, a link having a small weight is generated in the subgraph, and the degree of matching is calculated for each subgraph. 19. The storage device according to claim 18, further comprising a fourth calculation process of calculating a sum of the degrees of matching calculated for each graph to calculate a degree of matching between the graphs expressing the subject. Medium. 24. A storage medium storing a document classification program mounted on a document classification device for classifying documents, wherein a document corresponding to a document ID is obtained from a document storage unit in which the document is stored. Subject graph creation process for creating a graph representing the subject of the document, a graph similarity determination process for determining the degree of coincidence between the two graphs representing the subject of the two documents, A storage medium storing a document classification program, comprising: a classification process for classifying the document based on a matrix representing the similarity of the classification process; and a classification control process for controlling the entire classification work of the classification process. 25. The subject graph creation process includes the steps of: determining a degree of agreement between graphs representing the subject by combining word strings or Boolean operator combinations of word strings or sentences or documents or document collections; A word extraction process for extracting words used in the document element from the composed document elements; an importance assignment process for assigning importance to each of the extracted words; A relevancy assigning process for assigning a relevancy between two words, and expressing the subject of each document element by a graph in which the importance of the word is set as a node weight and the relevance between the words is set as a link weight. 3. A subject expressing process, and a document element similarity determining process for determining a similarity between the document elements based on a degree of coincidence between graphs expressing the subject. A storage medium storing the document classification program according to 4. 26. The graph similarity determination process, wherein the greater the number of similar nodes (nodes containing the same word) in both graphs, the greater the degree of coincidence between the graphs. If a node in the graph has a large weight, the greater the weight of a similar node in the other graph, the larger the degree of matching between both graphs, The greater the number of similar links (links having the same word in the nodes at both ends of the link) in both graphs, the greater the degree of matching between the two graphs, If a certain link is heavily weighted, the greater the weight of a similar link in the other graph, the greater the degree of matching between the two graphs, so that First storage medium storing a document classification program according to claim 24, further comprising a calculation process of calculating the degree of match between the graph representing the problem. 27. The graph similarity determination process includes dividing each graph into subgraphs based on how strongly the word sets used in the graph are related to each other; When there is no link between any nodes in the subgraph in each of the subgraphs, a link having a small weight is generated in the subgraph, the subgraph is recombined, and the subgraph is generated in the subgraph. Add the link as is,
25. The storage medium storing the document classification program according to claim 24, further comprising a second calculation process for returning the graph before division and calculating the degree of coincidence between the graphs expressing the subject. 28. The graph similarity determination process may include rewriting each graph based on how strongly the word sets used in the graph are related to each other.
If the subgraph is divided into subgraphs, and if each of the subgraphs has no link between any nodes between the subgraphs, a link having a small weight is generated in the subgraph; 25. The storage medium storing the document classification program according to claim 24, further comprising a third calculation process for calculating a degree of coincidence for each document. 29. The graph similarity determination process divides each graph into subgraphs based on how strongly the word sets used in the graphs are related, When there is no link between any nodes in the subgraph in each of the subgraphs, a link having a small weight is generated in the subgraph, and the degree of matching is calculated for each subgraph, 25. The document classification program according to claim 24, further comprising a fourth calculation process of calculating a degree of match between graphs expressing the subject by calculating a sum of degrees of match calculated for each subgraph. Storage medium.