JPH10116290A - Document classification managing method and document retrieving method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain a document classification managing method in which document contents are appropriately reflected or a document retrieving method which has a high efficiency even for a new document and a document including different classification systems. SOLUTION: Document parameters put together with tags of specified formats of N pieces of gathered documents are extracted. Then, the weight based upon the appearance frequency of components of an M-dimensional document parameter vector consisting of M pieces of document parameters extracted from the N documents is calculated to determine document parameter vectors by the documents. The similarities of document parameter vectors found by the documents are found, and the said gathered documents are classified and manages by document classes of the classification systems determined by the similarities.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

[0001] 1. Field of the Invention [0002] The present invention relates to a document classification management and document retrieval method for classifying and managing documents based on document information.

[0002]

2. Description of the Related Art When constructing a database by filing various documents, it is necessary to create information for retrieval such as classification and keyword assignment. In addition, since documents collected from a plurality of providers use different classification systems and terms depending on the providers, it is necessary to prepare different document search methods for each provider. In order to avoid this and search documents in a single framework, it is necessary to unify the document classification system and terms used in the documents. These tasks are important because their accuracy will lead to the convenience of subsequent use of the database, but if humans are to perform these tasks, they must read and understand the entire contents of the document. In addition, the work becomes extremely complicated, and the accuracy varies due to variations among operators.

Therefore, various systems have been developed as systems for automatically classifying documents. For example, as disclosed in Japanese Patent Application Laid-Open No. 6-282587, "a method and apparatus for automatically classifying documents and a method and apparatus for creating a dictionary for classification", the degree of field dependence is determined by keywords extracted while being classified from words constituting the document. Is calculated, and the field to which an arbitrary document belongs is determined from a known classification system.

However, in this method, since the classification field is determined based on the keywords extracted from the document, the exact term as the classification key in the document depends on the contingency of whether it is used in an identifiable notation. Therefore, it lacks reliability. In addition, the attribute given to the keyword in the above-mentioned “keyword extracted while being classified” is a grammatical case (nominative case or purpose case) and does not necessarily match the semantic case, so that the sentence creator such as active voice or passive voice is used. Lack of accuracy, such as depending on the wording of Furthermore, in the above-described classification method of “determining the field to which an arbitrary document belongs from a known classification system”, the document classification system information optimized by the provider is discarded, and a single known fixed classification is performed. Since it is incorporated into the system, it is not possible to cope with new fields or to refine the classification system with an increase in the number of documents.

[0005] As another method of document classification, for example, a keyword depending on the classification is extracted from the classified document data as in a "keyword assignment system" disclosed in Japanese Patent Laid-Open No. 7-78182, and the document is classified. There is a method of assigning keywords as keywords. However, even if the documents are similar, documents that do not include a specific keyword are omitted from the search.

As another method, for example, Japanese Patent Laid-Open No.
A plurality of document parameters are assigned to each document, as in the “attribute-based classification / search system” disclosed in Japanese Patent No. 282777, and classification is performed from a plurality of viewpoints using a single document parameter arbitrarily selected as a classification axis. There is a method of simultaneously configuring the system, but when the number of document parameters increases, the work of selecting a document parameter to be used as a classification axis becomes a burden on the searcher.

[0007]

A conventional document classification management method or search method is configured as described above, and a human intervenes to understand the contents of a document to determine and manage classification rules. Alternatively, the document is classified based on a fixed keyword, or a plurality of document parameters are assigned, a document parameter axis for classification is set, and the classification is performed based on the set document parameter axis. These methods are
There are problems such as the above-mentioned inconveniences, such as difficulty in work, leakage and lack of reliability, and large variation in results due to the responsibility of selecting a classification standard.

SUMMARY OF THE INVENTION The present invention has been made to solve the above-described problem. A document parameter is extracted from a document by using original classification information, and a document classification is performed using a document parameter vector composed of the document parameter. To obtain a document classification management method that appropriately reflects the document contents, or to obtain an efficient document search method for new documents and for documents that go beyond different classification systems. Aim.

[0009]

A document classification management method according to the present invention includes a document parameter extracting step of extracting document parameters of a plurality of collected N documents by tags of a predetermined format, and a step of extracting the document parameters. A document that determines a document parameter vector for each document by calculating a weight based on the appearance frequency of components of the parameter vector for an M-dimensional document parameter vector composed of M document parameters extracted from the document A similarity calculation step; and a document classification step of finding a similarity between the document parameter vectors obtained for each document and classifying the document into a document class of a classification system determined by the similarity. Documents are now classified and managed.

The document retrieval method according to the present invention includes a document parameter extracting step of extracting document parameters created by tags of a predetermined format from a plurality of collected N documents, and M extracted from the N documents. M composed of individual document parameters
A document similarity calculation step for calculating a weight based on the frequency of appearance of the component of this parameter vector for the two-dimensional document parameter vector, and determining a document parameter vector for each document. A document classification step of finding a similarity of the document and classifying the document into a document class of a classification system determined by the similarity; and a document conforming to the document class indicated in the classification system with respect to the plurality of classified and managed documents. Added a document classification search step to extract.

Further, when the appearance ratio of the specific document parameter to all the documents is specified for the documents classified and managed by the document classification management method, the document parameters having the predetermined appearance ratio close to the specified appearance ratio are added to the basic configuration. Added document parameter search guidance step to extract.

Further, an average parameter vector of the document class in the same classification system is obtained, and if the similarity of the obtained average parameter vector between the different classification systems is within a predetermined value, the merged classification combining different classification systems is performed. Added the step of merging the classification system to use the same document class in the system.

Further, in the classification system merging step, if there is a lower document class connected to the document class whose similarity is to be obtained, the lower connection form is managed as a merged classification system.

Further, a classification system subdivision step of extracting the most similar documents belonging to the same document class by calculation and sequentially moving them, and extracting the inter-document similarity for each movement is added.

Further, the terms and attributes extracted from the document by a predetermined term extraction rule are constructed and managed in correspondence with the document class as a term attribute dictionary, and when a specific attribute is designated, the attribute corresponding to this attribute is specified. Added a document classification search guidance step to extract the document class that has the term.

Furthermore, a keyword search mechanism for constructing and managing terms and attributes extracted from a document according to a predetermined term extraction rule in association with a document class as a term attribute dictionary and searching for a document having a specific keyword is provided. In addition, when a specific attribute is specified, the corresponding term is extracted and searched as a keyword.

Further, the terms and attributes extracted from the document by a predetermined term extraction rule are constructed and managed as a term attribute dictionary corresponding to the document class, and when a specific attribute is designated, the term corresponding to the attribute is designated. And a similar document adding step of extracting and adding a predetermined number of documents in descending order of similarity with the search result document class.

Further, a synonym / similar word dictionary is constructed and managed based on the similarity between terms having the same document class attribute,
A keyword search mechanism that searches for documents with specific keywords has been added.
A keyword search is performed in addition to the keyword registered and managed as equivalent in the similar word dictionary.

Still further, a characteristic term determining step of extracting characteristic terms based on the frequency of appearance in a specific document class;
A keyword search mechanism for searching for a document having a specific keyword is added, and in the keyword search, the characteristic term extracted in the characteristic term determination step is searched as the keyword.

[0020]

BEST MODE FOR CARRYING OUT THE INVENTION

Embodiment 1 FIG. A first embodiment of a document classification management method and a document search method according to the present invention will be described in detail with reference to the accompanying drawings. FIG. 1 is an operation flowchart of a basic document search method according to the present invention. FIG. 2 is a diagram showing a document providing source for explaining the operation of FIG. 1, an example of a specific document to be analyzed, and an example of extracted “characteristics”. FIG. 3 is a diagram for explaining the operation of the document similarity calculation.
FIG. 3A is a diagram illustrating an example of a document whose characteristics to be analyzed are tagged and a document parameter, FIG. 3B is a diagram illustrating an example of a document parameter vector, and FIG. 3C is a diagram illustrating an example of a classification system. .
FIG. 4 is a flowchart of the document similarity calculation operation.

First, as shown in FIG. 1, documents of a plurality of providers are collected by a document collecting step 1. In the document collecting step 1, the location information of the document of the provider stored in the form of the address on the network as shown by the source address 1001 in FIG. 2A is read, and the document of the provider is read by a known file transfer protocol. Download using such as.

Next, document parameters are extracted from each document collected in the document collection step 1 in a document parameter extraction step 2. The document parameter is information on “characteristics” of the thing described by the document. This "characteristic"
The information is already indicated in the document to be analyzed by pairing “characteristic name” and “characteristic value” with a tag. That is, each document is SGML (Structured Document Description Language: I
SO8879), and in document parameter extraction step 2, SGML
The document parameters are extracted by analyzing the format of the document content using a known SGML parser or the like that checks the format of the document. For example, assuming that the provider is a component maker and the document is a component specification, the "characteristics" of the component are described in the document content by enclosing the tag as shown in FIG. The “characteristic name” (or simply “name”) and the “characteristic value” of the part are extracted as document parameters. If the document is a specification of a memory chip, tags <parts specification>, <characteristics>, <name>, <value>, </ parts specification> described in the document A in the SGML format shown in FIG. Is analyzed. As a result, the “name” of the document parameter such as the memory capacity, refresh cycle, and access time as shown in FIG.
B, “characteristic values” such as 20 ns and 10 ns are extracted.
In the example of FIG. 2A, documents B and C from vendors B and C are used.
Similarly, the document parameters are extracted from, and the extraction result of FIG. 2C is obtained.

The product catalog 101 shown in FIG.
0 is an example of a product catalog of the display monitor described in the SGML format. Tags in this catalog: <product specification>, <product characteristics>, <characteristic name>, <characteristic value>, <//
By analyzing the product characteristics> and </ product specifications> in the same manner, the “name” of the document parameter “CRT size” shown in the document parameter 1011 in FIG. 3A and the “characteristic” of the 17-inch document parameter corresponding thereto are displayed. "Value", the document parameter "name" of dot pitch and the corresponding "characteristic value" of the document parameter of 0.25 mm, the horizontal scanning frequency of the document parameter "name" and the corresponding document parameter of 24 kHz to 86 kHz "24 kHz to 86 kHz". The characteristic value is extracted as a document parameter.

Next, the similarity between documents is calculated in a document similarity calculation step 3 using the document parameters extracted in the document parameter extraction step 2 of FIG. The document similarity calculating step 3 creates a document parameter vector as shown in FIG. 3B for each document according to the flowchart shown in FIG. The result of calculating the inner product between document parameter vectors as described later is defined as the similarity between documents.
The document parameter vector is a vector in which each document parameter is associated with each vector component. Assuming that N documents are collected and M types of document parameters are extracted, the document parameter vector is a vector including M-dimensional components. Numbers i of 1 to M are assigned to M types of document parameters, and the i-th document parameter P (i) corresponds to the i-th document parameter vector component. At this time, the document parameters are distinguished by the names of the document parameters, and the document parameters having the same document parameter name are regarded as the same type of document parameter. Numbers 1 to N are assigned to the N documents, and the i-th document parameter vector component of the j-th document D (j) is V (j, i). However, it is necessary
When the (M + 1) th parameter vector component is provided, the M + 1 component is set to 0.

In FIG. 4, steps S100 to S10
5 is a process for calculating the weight C (i) of each component of the document parameter vector based on the appearance frequency of the document parameter. In step S100, the document parameter number i is initialized to 1, and
In step S102, the i-th document parameter is set to P (i)
Set to. In step S101, it is checked whether or not the document parameter number i exceeds the maximum value M, and if so, the loop of steps S101 to S105 ends. In step S103, the number of documents having the i-th document parameter P (i) is counted out of the N documents, and the value is set in S. In step S104, i
The weight C (i) of the document parameter vector component of the
Set to S. In step S105, i is increased by one, and the weight of the document parameter vector component corresponding to the next document parameter is calculated. When the value of i exceeds M, the weights C (1) to C (M) of all components of the document parameter vector
Is completed. For example, in the three documents (N = 3) shown in FIG. 2C, the document having the second (i = 2) document parameter P (2) “refresh cycle” includes two documents A and B. (S = 2), the weight C (2) of the second document parameter P (2) is P (2)
Ｓ, which is the reciprocal of the number of documents having S (2).

Steps S109 to S114 are processes for calculating the temporary document parameter vector of the j-th document D (j) and the square L of its length. In step S109, the square L of the length of the temporary document parameter vector is initialized to zero.
In S110, the document parameter number i is initialized to 1. S1
At 11, it is checked whether or not the document parameter number i exceeds the maximum value M, and if it does, the loop of S111 to S114 ends. In S112, the j-th document D (j) is i
Check if it has the th document parameter P (i). The j-th document D (j) is the i-th document parameter P
If it has (i), the i-th component of the temporary document parameter vector of D (j) is set to C (i).
Add the square of (i). If the j-th document D (j) does not have the i-th document parameter P (i), ie,
For the (M + 1) th component, the ith component of the temporary document parameter vector of D (j) is set to 0. I in S114
Is increased by one, and the value of the temporary document parameter vector component of D (j) corresponding to the next document parameter is calculated. When the value of i exceeds M, the calculation of the temporary document parameter vector component of D (j) and the square L of its length is completed.

Steps S106, S107, S115-
S122 is processing for calculating document parameter vectors of all N documents. In step S122, the document number j is initialized to 1, and in step S107, the j-th document is set to D (j). In S106, it is checked whether or not the document number j exceeds the maximum value N, and if it does, S106, S10
7. The loop of S109 to S121 ends. S115
Initializes the document parameter number i to 1. In S116, it is checked whether or not the document parameter number i exceeds the maximum value M, and if so, the loop of S116 to S120 is ended. In S117, it is checked whether the j-th document D (j) has the i-th document parameter P (i). j
If the document D (j) has the i-th document parameter P (i), the temporary document parameter vector i
The value C (i) is the length of the temporary document parameter vector √
The value divided by L is used as the value of the document parameter vector as V
Set to (j, i). If the j-th document D (j) does not have the i-th document parameter P (i), 0 is set to V (j, i) as the value of the document parameter vector in S119. In S120, i is increased by one, and the value of the document parameter vector component of D (j) corresponding to the next document parameter is calculated. When the value of i exceeds M, the calculation of the value of the document parameter vector component of D (j) ends. S121
J is incremented by one, and the document parameter vector of the next document is calculated. When the value of j exceeds N, the calculation of the document parameter vectors of all the documents is completed.

For example, the third (j = 3) shown in FIG.
In document C, the temporary document parameter vector is (C
(1), 0, C (3), 0) = (1/3, 0, 1/3,
0), that is, since there is no fourth characteristic value, the fourth component is 0, and the square of its length L is 1/9 + 1 /
9 = 2/9. By dividing each component value V (3, i) of the temporary document parameter vector by its length ΔL, the document parameter vector (V (3,1), V (3,2), V (3,3), V
(3, 4)) = (0.71, 0, 0.71, 0) is obtained. For document A and document B, each component value of the document parameter vector is calculated in the same manner, and a document parameter vector as shown in FIG. 3B is obtained for each document.

The documents for which the similarity between the documents can be calculated by the document parameter extraction step 2 and the document similarity calculation step 3 are the most similarity in the classification system separately determined by the document classification step 4 in FIG. To a document class with high Here, the similarity between the document D (m) and the document D (n) can be calculated as follows: Inner product between document parameters = V (m, 1) × V (n,
1) + V (m, 2) × V (n, 2) +. . . + V (m,
M) × V (n, M). Then, the same document class or more is set as the same document class. For example, assuming that a separately defined classification system is given as shown in FIG. 3C, in the document classification step 4, the similarity is determined by the document parameter vector of the collected document and the document parameter vector given to the document class. Ask for. Document A is shown in FIG.
The similarity with the document class DRAM 1005 in FIG.
(B) Document parameter vector 1004 in FIG. 3 and FIG.
By calculating the inner product with the document parameter vector 1007 in (c), 0.986 is obtained. Similarly in FIG.
The similarity with the document parameter vector of the document class SRAM 1006 in (c) is obtained as 0.616. Similarly, the document parameter vector similarity between all the document classes and the document A is calculated. As a result, when the similarity between the document A and the document class DRAM 1005 in FIG. Document A has the highest degree
Is the document class DRAM having the highest similarity in FIG.
Register in 1005. As described above, it is possible to determine in which document class of the existing classification system the collected documents are to be registered, so that the work in document registration is facilitated. By the above operation steps, a plurality of collected documents can be classified and managed.

The classified documents are searched by selecting a document class in the classification system in the document classification search step 5 of FIG. For example, a classification system as shown in FIG. 3C is presented to the searcher. When the searcher selects the "memory" of the document class, the document classification search step 5 is performed in the document class "DRAM" or the document class "SR".
Only the documents registered in "AM" are presented to the searcher. Next, when the searcher selects the document class “DRAM”, the document classification search step 5 presents only the documents registered in the document class “DRAM” to the searcher. The search function can be realized by a well-known relational database search function using the document class name as an attribute.

Embodiment 2 FIG. 5 is an operation flowchart showing a search method according to the second embodiment of the present invention. In the drawing, the same or corresponding parts as those in the first embodiment are denoted by the same reference numerals and description thereof is omitted. In the present embodiment, in the document classification management method in the first embodiment,
At the time of retrieval, the expected appearance ratio of a document having a specific document parameter in all documents is defined as a search cost, and when this search cost (expected appearance ratio) is specified, a step of extracting a document parameter close to the search cost is added. . By doing so, the search range can be narrowed more easily.

Document parameter retrieval step 6 in FIG.
Searches for a document using the document parameter described in the first embodiment as a search condition. This search function can be realized by a known method by constructing a relational database using document parameters as attributes and attribute values.

The document search method according to the present embodiment is characterized in that the search condition input in the document parameter search step 6 is performed using the search guidance created in the classification parameter search guidance step 7. The search guidance created in the classification parameter search guidance step 7 evaluates the document parameters of the document to be searched from the viewpoint of search efficiency, and displays the document parameters in order of the search efficiency. Alternatively, the number of cases for each “name” or “characteristic value” with respect to the number of target documents is displayed. This has the effect of prompting the searcher to input a document parameter with high search efficiency as a search condition. The search target document is a set of documents in the middle of the search when the documents are narrowed down by repeating the document search a plurality of times. Therefore, the classification parameter search guidance step 7 is executed each time the document search is repeated.

FIG. 6 is a flowchart showing the details of the operation of the classification parameter search guidance step 7. Here, the type of the document parameter of the document to be searched is I, the number assigned to the document parameter of the document to be searched is i, and P (i) is the i-th document parameter.
The number of documents to be searched is J, the number assigned to the documents to be searched is j, and D (j) is the j-th document.
Here, i: the number of the document parameter, I: the type of the parameter of the search target document, P (i): the i-th document parameter, j: the number of the search target document, J: the number of the search target document, D ( j): j-th document, S (i): number of documents having P (i), Q (i): search cost of P (i), B (i): S (i) ≠ J, C1: search The optimal value of the document ratio, C2: the minimum value of the document parameter value type, C3: the minimum ratio of the real number distribution, and C4: the document parameter value division number. Steps S202 to S208 in the upper left loop mean counting the number of documents having a certain document parameter value.
Steps 214 to S221 mean an operation of dividing a target document having a common parameter described later and targeting the divided document. Step S212 in the middle right shows the cost calculation of the target document having a unique parameter described later, that is, the appearance frequency. The position is determined.

In steps S201 to S208, P (i)
Is calculated. Step S201
Then, 0 is set as an initial value in S (1) to S (I), and 1 is set as an initial value of document number j. J in S202
Does not exceed the number J of documents to be searched, and if it does, the calculation of S (i) ends. In S203, 1 is set as the initial value of the document parameter number i, and in S20
In step 4, i is the document parameter type I of the document to be searched.
Is checked, and if so, the document number j is incremented by one in S205 and the next document is processed. S20
6 checks whether D (j) has P (i),
If so, S (i) is increased by one. I in S208
Is incremented by one and the next document parameter is processed. As described above, the number S (i) of documents having each document parameter is calculated for all document parameters of the document to be searched.

Steps S209 to S221 are processes for evaluating the document parameters of the document to be searched from the viewpoint of search efficiency and calculating the search cost Q (i). The search cost here is an expected appearance ratio of the ratio of the target document to be extracted with respect to all the documents. Logically 1
/ 2 is the smallest number of times to narrow down the target. Here, it is assumed that the smaller the search cost Q (i), the higher the search efficiency of the document parameter (i). S209, S210, S21
3, the document parameter number i is changed from 1 to I, and the search cost Q (i) of all document parameters is calculated.
The number S of documents having the document parameter P (i) in S211
It is checked whether or not (i) matches the number J of documents.

The case where S (i) and J match, that is, when all documents to be searched have document parameters P (i), such document parameters P (i) are called common parameters. "Name" for common parameter P (i)
In this case, it is necessary to use the “characteristic value” as a search condition because the target cannot be narrowed down. Therefore, the common parameter P (i)
Q (i) is calculated in steps after S214 based on the distribution of the values of. When S (i) and J do not match, that is, when only a part of the document to be searched has the document parameter P (i), such a document parameter P (i) is called a unique parameter. Since the existence of the unique parameter P (i) needs to be used as a search condition, the unique parameter P (i)
Based on the existence ratio of (i), Q (i) is calculated in step S212.

In S212, the smaller of the absolute value of J * C1-S (i) or the absolute value of J * (1-C1) -S (i) is set as the search cost in Q (i). For example, when the number J of documents to be searched is 120 and the optimum value C1 of the search document ratio is set to 1/3, S (i) is 40 or 80, which is a multiple thereof, is 1/3 and the remainder is 1/3.
(I) is the document parameter with the lowest search cost, that is, the document parameter with the highest search efficiency. In S212, P
Save that (i) is a unique parameter by setting B (i) to one.

In steps S214 to S221, the search cost Q (i) of the common parameter P (i) is calculated from the distribution of the value of the common parameter P (i) commonly for the documents to be searched. In S214, the number of values of the common parameter P (i) for the J documents is counted and set to X. In step S215, it is checked whether the type X of the value of the common parameter P (i) is larger than the minimum value C2 of the type of the document parameter value set in advance. If X is equal to or less than C2, the value of the common parameter is considered to be discrete, and S2
At 20, the number of identical parameter values is counted, and the maximum value is set to M.

When X is larger than C2, the value of the common parameter is considered to be continuous, so that the range of the common parameter value is treated discretely by dividing it into a predetermined number of divisions C4. At this time, the ratio W between the maximum value of the absolute value of the document parameter value and the minimum value of the absolute value is determined in S216.
Is checked in step S217 to determine whether W is larger than a predetermined distribution ratio C3. If W is greater than C3, the parameter value is divided into C4 pieces at equal intervals on the logarithmic axis in S219, and if W is less than C3, the parameter value is divided into C4 pieces at equal intervals on the real axis in S218. As a result, the parameter value is set to S220 in the same manner as when X is C2 or less.
Can be treated discretely. In S221, S21
2, the absolute value of J × C1-M or J × (1-C
1) The smaller one of the absolute values of -M is set as a search cost in Q (i). In S221, P (i)
Is a common parameter by setting B (i) to 0. By repeating the above processing in step S210 until the document parameter number i exceeds the document parameter type I, it is determined whether all document parameters are common parameters or unique parameters B (i), and the search cost Q ( i) is calculated.

In step S222, the document parameter P (i) is divided into two types, a common parameter and a unique parameter, according to the value of B (i). In step S223, the divided document parameter P (i) is divided into search costs Q ( The items are ranked in ascending order according to i). Further, in step S224, the co-occurrence parameter is removed from the unique parameter P (i). A co-occurrence parameter means that all documents having A in the unique parameter A and the unique parameter B are B
, And all of the documents having B are combinations of the unique parameters A and B having A. Unique parameters other than any one of the unique parameter combinations in the co-occurrence relationship are deleted in S224, and the unique parameters of lower ranks are moved up in rank.

It is also possible to display the count number for each document parameter of the extracted document of the search result without specifying the search cost. FIG. 7 shows an example of a document parameter search condition input screen in the classification parameter search guidance step 7. The common parameter name display 37 displays the names of a fixed number of common parameters from the top of the ranked common parameters. Parameter value input area 38 paired with common parameter name display 37
, A search condition based on the value of each common parameter is input, and a search is executed by a search execution button 43. In this embodiment, if the new button 40 is selected, the search target document is discarded and the entire document is set as the search target so that the search range can be further changed. Add button 41
If is selected, the search result for the entire document as a search target is added to the search target document, and if the narrow down button 42 is selected, the current search target document is set as the search target. The state before the search is restored by the return button 44 after the search is executed.

The number of documents currently being searched is displayed based on the number of documents to be searched 39. This example indicates that there are 207 documents including the common parameter “name”. FIG. 7B shows the number of cases for each “characteristic” value with respect to the “name” minimum operating power supply voltage of the common parameter, and the input value candidate and each input value are input by the input candidate button 49 in the parameter value input area 38. A common parameter input value candidate 50 including the number of documents belonging to the candidate is displayed. For example, here, it can be seen that there are 56 documents having a minimum operating power supply voltage value of 3.6 V to 4.0 V in the search target document. Further, a common parameter value distribution graph 5 is displayed by a distribution button 45.
1 is displayed, and the value of the common parameter and the distribution state of the number of documents are graphically displayed.

The unique parameter name 47 indicates the name of a fixed number of unique parameters from the top of the ranked unique parameters. The presence / absence of a unique parameter is input as a search condition by a parameter selection button 46 paired with the unique parameter name 47. The number 48 of unique parameter-containing documents forming a pair with the unique parameter name 47 indicates the number of documents including the unique parameter. The system selects three types of selection states of the parameter selection button 46, and sets a document search condition that includes the specified unique parameter, a document search condition that does not include the specified unique parameter, and a search condition that specifies the specified unique parameter. It is also possible to correspond to three document search conditions of not performing. As described above, in the parameter search that searches for documents by document parameters, guidance for inputting search conditions is generated by analyzing the document to be searched, and presented to the searcher. This has the effect of making it easier to input.

Embodiment 3 FIG. 8 is an operation flowchart of the document search method according to the present embodiment. The same or corresponding parts as those in the first embodiment are denoted by the same reference numerals and description thereof is omitted. The search method according to the present embodiment is applied to the documents classified and managed in the first embodiment.
Documents under different classification systems are integrated and managed beyond the classification system, or integrated and searched. By doing so, the search range can be expanded.

In the drawing, a classification system acquisition step 22 simultaneously collects the classification system constructed by the provider when collecting documents in the document collection step 1. A table of contents that refers to a set of documents exists at the provider, and the contents of the table of contents are SGML.
If the document is described in a predetermined format such as (Structured Document Description Language: ISO8879), the document is analyzed by analyzing the structure of the table of contents using a known SGML parser or the like for checking the format of the document in the SGML format. Classification system can be obtained.

In a classification system merging step 8, a plurality of classification systems obtained in the classification system acquisition step 22 are merged,
This is a process for constructing a single unified classification system. FIG. 9 is a flowchart for explaining the details of the operation of the classification system merging step 8. FIG. 10 is an explanatory diagram of the operation.
In FIG. 10, two classification systems T (1) and T (2) are merged. First, in FIG. 9, i: classification system number (i = 1 to 2), T (i): i-th classification system, J (i): number of terminal document classes of T (i), j: terminal Document class number, C (i, j): j-th terminal document class of T (i), V (i, j): average parameter vector of C (i, j), K (j): V (1) , J) and V (2, k) have the maximum similarity. Step S3 of the upper left loop of this operation flow
01 to S307 mean finding the average vector V (i, j) of a certain document class, and steps S308 to S314 in the right loop calculate the similarity of the average vector between document classes having different classification systems. This means that steps S315 to S32 of the lower left loop
0 means the operation of merging while maintaining the connection relation of the lower document classes even after the merging of the classification systems.

By the steps S301 to S307 in FIG. 9, all the document classes C located at the end of T (i)
The average vector V (i, j) of (i, j) is calculated. S
The number i of the classification system T is changed from 1 to 2 by 301, S302, and S305, that is, the classification system T (1) is changed from T (2), and the end document class number is changed from 1 by S303, S304, and S305. J (i), and by calculating the average vector in S306, all the document classes C of the classification systems T (1) and T (2) are obtained.
The average vector V (i, j) of (i, j) is calculated. Average vector V (i, j) is document class C
In the document parameter vectors of all the documents belonging to (i, j), a temporary average vector based on the average value of each document parameter vector component is created, and the document parameter vector is normalized to have a length of 1.

Next, in steps S308 to S320, the classification system T (2) is set as a reference classification system, and all the document classes of the classification system T (1) have the highest similarity to any document class of the classification system T (2). Is determined. S30
8, S309 and S312 change j from 1 to J (1), and process all the terminal document classes C (1, j) of the classification system T (1), and at the same time, S310, S31
1, k is changed from 1 to J (2) by S314,
All terminal document classes C (2,2) of the classification system T (2)
Process k). At this time, V is set for j in S313.
The degree of similarity between (1, j) and V (2, k), that is, k that maximizes the inner product value, is determined, and this is determined as the maximum similar document class number K
Set to (j). That is, C (2,1) to C (2,1) which are all terminal document classes of the classification system T (2).
J (2)), the terminal document class most similar to C (1, j) is C (2, K (j)).

In S315 to S320, the classification system T (1)
Is performed to combine each end document class C (1, m) into the classification system T (2). S315, S316, S320
Changes m from 1 to J (1) according to the classification system T
Process all terminal document classes C (1, m) in (1). In S317, n (n) where K (m) = K (n) is satisfied.
≠ m) is determined to be present or not. K
If (m) = K (n), the two or more end document classes C (1, m) and C (1, n) of T (1) are all T
It has the highest similarity to the last document class C (2, K (m)) of (2). In this case, C is determined by S319.
(1, m) and C (1, n) are combined assuming that they are lower-order document classes of C (2, K (m)). At this time, C (2,
K (m)) belong to C (1, m), C
The process moves to the document class having the higher similarity of (1, n). If n where K (m) = K (n) does not exist, C (1, m) is changed to C (2, K
(M)) and combined assuming that they are the same document class.

FIG. 10 is a diagram illustrating the above operation. In the figure, the output device 903 of the terminal document class of the classification system T (2) has the highest similarity with the output device 901 of the terminal document class of the classification system T (1).
Since there is no other terminal document class of T (2) whose output device 901 of the terminal document class of T (1) has the highest similarity, the output device 9 of the terminal document class of T (2) exists.
03 is an output device 901 of the terminal document class of T (1)
The document is regarded as having the same document class and is combined, and a new output device 906 of the terminal document class is created. T
Since both the hard disk 904 and the FDD 905 of the terminal document class of (2) have the highest similarity with the storage device 902 of the terminal document class of T (1), the hard disk 904 and the FDD 905 of the terminal document class of T (2).
Are both storage devices 902 of the terminal document class of T (1)
It is considered as a lower document class of and combined. As described above, the document classification system is collected from the provider at the same time as the document, and the similarity between the document classes is calculated. With this, documents can be merged beyond the classification system, classified, managed, and registered. Therefore, in a subsequent search, documents having a high degree of similarity can be handled as the same document class. If the classification system of the document collected from the provider is more detailed than the existing classification system, it is possible to refine the existing classification system.

Embodiment 4 FIG. FIG. 11 is an operation flowchart of the document search method according to the present embodiment, and the same or corresponding parts as those in the third embodiment shown in FIG. Also, the classification system merging step 8 is not essential. According to the present embodiment, it is possible to improve a decrease in search efficiency, such as a case where the number of documents belonging to the same document class increases and a search takes time, and to manage documents of the same class by subdividing the search range. .

The classification system subdividing step 9 subdivides the classification system merged in the classification system merging step 8. FIG. 12 is a flowchart for explaining the flow of the operation of the classification system subdividing step 9. In the figure, S: move step, i, j: number of documents in a specific document class, I: number of documents in a specific document class, J: number of destination document, C (i): specific number I-th document in the document class, Sim (A, B): similarity between document A and document B, P (i): passage flag of document C (i), L (s): movement of movement step S The similarity between documents, Steps S401 to S412 are processing to be started when conditions such as the number of documents belonging to a certain terminal document class exceed an upper limit value and to create a lower-level document class of the terminal document class. The classification system is subdivided by applying to the end document class of.

Steps S401 to S404 are initial settings. S
At 401, the pass flags P (i) of all documents are reset to zero. By setting 1 to the document number i in S402, the document C (1), which is an arbitrary one of the documents belonging to the specific end document class, is selected as the starting point of the moving document. 1 is set as the initial value of the moving step s. In S403, 2 is set as the initial value of the destination document number j for searching for the destination document. S4
In step 04, the maximum value M of the similarity between the moving documents is initialized to 0, and the pass flag P (i) of the moving source document is set to 1. The destination document number J is determined by the repetition of steps S405 and S410.

Since the destination document must be a document that has not yet passed, the destination document candidate C is determined in step S406.
The pass flag P (j) of (j) is checked, and if P (j) is 0, the similarity between the source document C (i) and the destination document candidate C (j) is calculated in S407 and S Set to. S
At 408, it is checked whether or not the similarity S between the moving documents exceeds the maximum value M, and if it does, the maximum value M of the similarity between the moving documents and the document number J of the moving destination are updated at S409. If it is determined in step S405 that all the investigations of the candidate j of the destination document number have been completed, the destination document number J and the maximum value M of the similarity between the moved documents are determined. That is, in steps S405 to S410, it means that all the most similar documents having the highest similarity with the first selected document C (1) are compared with all the remaining documents in the same document class, and selected.

If the maximum value M of the similarity between moving documents is 0 in S411, it is determined that all the movements between the documents have been completed, and the process is terminated. If the maximum value M of the similarity between moving documents is not 0, the document is moved from the document C (i) to the document C (J) in S412, and the similarity L (s) between the moving documents at that time is set to M. I do. The document is moved in the next step by increasing s by one and setting i to J. As described above, in a specific document class, an arbitrary document in the document class is set as a starting point, a document that has passed once is not passed twice, the document is successively moved to a document having the highest similarity, and all documents are moved. I do. At this time, by storing the similarity L (s) between the moving documents in the moving step s, the relationship between the moving step s and the similarity L (s) between the moving documents is obtained.

FIG. 13A illustrates the above processing. A document (7) having the highest similarity in the same document class, starting from an arbitrary document (1) 121 in the document class
Move to 122. The document that has passed once is regarded as not passing again, and the movement to the most similar document is repeated in the same manner until all documents in the document class have passed.

As described above, the relationship between the document moving steps s and the similarity L (s) between the moving documents in each moving step as shown in FIG. 13B is obtained. That is, the similarity L between the moving documents corresponding to the moving step S in step S412 of the operation flow of FIG. 12 is sequentially plotted. In this way, generally, a certain moving step range W is set in advance, and the sum of a plurality of similarities L (s) between moving documents in the range from the moving step s to s + W is calculated. A plurality of maximal ranges 124 and 125 in which the sum of the similarities L (s) between the moving documents in the range up to is maximal are obtained.

By examining the document parameters of the document at each central portion of the plurality of maximum ranges from the plot obtained above,
It is easy to create a lower document class corresponding to each maximal range. That is, the moving step s for each maximum range
, The average vector of the documents belonging to the range of s + W is calculated, and this is used as the document parameter vector of the lower document class. Or, newly set the document parameters of the lower document class based on the result of the check. Documents belonging outside the maximum range are moved to any of the lower-order document classes having the highest similarity, and the document class can be subdivided. In the example of FIG. 12, at least two lower-order document classes are created by dividing a document group belonging to two groups of the maximum ranges 124 and 125. As described above, since a specific single document class is subdivided into a plurality of lower document classes, it becomes possible to subdivide the classification system of the document and manage the classification by subdividing the document system as the number of registered documents increases.

Embodiment 5 FIG. FIG. 14 is an operation flowchart of the document search method according to the present embodiment. The same reference numerals are given to the same or corresponding parts as in the first embodiment, and the description thereof will be omitted. The search method according to the present embodiment
Separately, a term paired with an attribute is extracted to create a term attribute dictionary, and by specifying a certain attribute, a document class including the attribute is extracted to facilitate retrieval. In the figure, a term extraction step 10 extracts terms from a text data portion in a document. The term extraction function can be realized by dividing text data into words by a known morphological analysis process and extracting words to which specific parts of speech such as nouns are attached.

The term attribute judging step 11 judges term attributes (or simply attributes) of the extracted terms by applying the term extraction rule to the text data divided in word units in the term extracting step 10. FIG. 15 is a flowchart for explaining the detailed operations of the term attribute determination step 11 and the term attribute dictionary construction step 12.
The term extraction rule R (i) in the figure includes a text pattern part and a term attribute extraction part as shown in FIG. 16A, and a part of text data is matched with the text pattern part of the term extraction rule. If successful, the term attribute extraction unit is executed to determine the attribute of the term. For example, FIG.
FIG. 14 shows the text portion of the document as shown in FIG.
When the term attribute determination step 11 of FIG.
A term with a term attribute as shown in (c) is extracted.
The word “for” included in the text portion of the document shown in FIG. 16B is successfully matched with the text pattern portion of the term extraction rule of rule number 1 among the term extraction rules shown in FIG. Since the word “answering machine” matches the variable # 1 of the pattern portion, the term attribute extracting unit determines that the term attribute of the word “answering machine” is “use”.

The terms for which the attributes have been determined are registered by adding term attribute information to the attribute dictionary in the term attribute dictionary construction step 12 of FIG. Term attribute judgment step 1
FIG. 15 showing details of step 1 and the term attribute dictionary construction step 12.
The operation will be described based on. In the figure, i: term extraction rule number, I: number of term extraction rules, R (i): i-th term extraction rule, the steps S502 to S508 are the text data divided in word units in the term extraction step 10. This is a process for applying all the term extraction rules to. S50
In step 1, a text description portion is determined by analyzing the format of a document. For example, if the document is in the SGML format, <
A range specified by tags such as “Overview”, “Description”, and “Product Description” is determined as a text description portion. S5
02, S503, and S508, the rule number i is changed from 1 to I, and all term extraction rules R (i) are applied in S504.

In S505, the text pattern portion of the term extraction rule R (i) is matched with the text pattern portion determined in S501, and if the matching is successful, the term attribute extraction portion of the term extraction rule R (i) is checked in S506. The attribute is determined by executing. In step S507, the term whose attribute is determined is registered by adding attribute information to the term attribute dictionary. Document classification search guidance step 13
Using the attributed terms extracted from the document as described above, the function of displaying the attributed terms of the documents belonging to each document class in the document classification search step 5 and the document class including the designated attributed term And a function of extracting and displaying.

It is assumed that the result shown in FIG. 17A is obtained by the term attribute determining step 11 and the term attribute dictionary constructing step 12, for example. Where document class T
Term / attribute 1502 is extracted from document 1501 belonging to (4). FIG. 17B shows a search screen in the document classification search step 5. Here, document class T
A classification system 1506 made up of (1) to T (5) is displayed.
When 505 is selected, the term list 1 of the attribute specified in 1503 among the terms of the documents belonging to each document class
507 is displayed. When a term is specified from the term list 1507 and a search button 1505 is selected, a document class having the specified term is displayed. For example, when “use” is selected as the attribute, the term list 1507 includes “answering machine”, “cooked rice” with the attribute “use” among the terms extracted from the documents registered in the document class T (4). Only terms such as "instrument" are displayed. As described above, the searcher can grasp the classification system from a semantic viewpoint by listing the document classes by the terms of the designated attribute. Further, by using the displayed term as a search condition, it becomes easy to input a search condition in a keyword search described later.

Embodiment 6 FIG. FIG. 18 is an operation flowchart of a document search method according to the present embodiment. The same or corresponding parts as those in the fifth embodiment are denoted by the same reference numerals, and description thereof is omitted. In the figure, a document full-text search step 14 is a well-known search process in which a document containing a keyword is set as a search result using a freely input keyword as a search condition.

The document full-text search guidance step 15 presents the attributed term extracted in the term attribute determination step 11 as an input candidate of a search condition keyword in the document full-text search step 14. FIG. 19 shows an example of a search condition input screen in the document full-text search step 14. When a search keyword is input to the search keyword input 1801 and a search button 1802 is selected, a document including the input keyword is searched. The document full-text search guidance step 15 adds a term attribute selection mechanism 1803 and a keyword candidate display mechanism 1804 to this. When an attribute is selected by the term attribute selection mechanism 1803, a keyword candidate display mechanism 1804 displays a term list of a document to be searched having the selected attribute. When a term displayed on the keyword candidate display mechanism 1804 is selected, the selected term is added to the search keyword input 1801 as a keyword for search. As described above, by inputting the search condition of the keyword search using the term of the term attribute designated by the searcher among the terms extracted from the search target document, the search condition input operation of the searcher becomes easy. Furthermore, since the terms extracted from the document to be searched are used as search conditions for the keyword search, at least one search result document is included.
Are guaranteed to exist.

Embodiment 7 FIG. 20 is an operation flowchart of a document search method according to the present embodiment. The same or corresponding parts as those in the sixth embodiment are denoted by the same reference numerals and description thereof is omitted. Also, as in the other embodiments, some steps may be omitted or combined.
The synonym / similar word judging step 16 extracts synonyms / similar words by judging the synonym / similarity of a plurality of terms having different notations extracted from a plurality of different documents in the term extracting step 10. Synonym / similar word dictionary construction step 17 is a synonym / similar word determination step 16
Constructs a synonym / similar word dictionary composed of the synonyms / similar words extracted in (1).

FIG. 21 shows a synonym / similar word determination step 16
6 is a flowchart illustrating the detailed operation of synonym / similar word dictionary construction step 17. In the figure, all document classes in the classification system are processed in steps S1901, S1902, and S1905. S1903, S
By 1904 and S1908, all documents belonging to a specific document class are processed. S1906, S1907,
In step S1915, all term attributes of the specific document are processed. In step S1909, a set W of terms to which a specific attribute A is assigned is extracted from the specific document D. S1
At step 910, a document E whose similarity with the document to be processed is equal to or greater than a predetermined lower limit L is obtained. S1911
, A set X of terms to which the attribute A is assigned is extracted from the document E. In S1912, a term set Y obtained by deleting terms common to the term set X from the term set W is obtained.
A term set Z obtained by deleting terms common to the term set W from the term set X is obtained. According to S1913, the term belonging to the term set Y is one Y (1), and the term set Z
Is Z (1), the term Y
(1) The term Z (1) is extracted as a synonym / similar word. At this time, the similarity between the document D and the document E is expressed by the term Y
The similarity between (1) and the term Z (1) is assumed. In S1914, the synonym / similar words Y (1) and Z (1) are registered in the synonym / similar word dictionary together with the similarity.

For example, if two documents registered in the same document class
There are two integrated circuit component specifications, and the similarity is 0.
Suppose the value was as high as 97. Two terms “cellular phone” and “cellular phone” are extracted as terms whose term attributes are “use” from the specification of one part, and “term” is a term whose term attribute is “use” from the specification of the other part. It is assumed that two items, "car phone" and "cellular phone", are extracted. The term "cellular phone" is a common term and is ignored, and the two remaining terms extracted from each document, one by one, "mobile phone" and "car phone" are synonyms.
Judge as a similar word, and set its similarity to 0.97,
Register in the similar word dictionary.

The document full-text search synonym / similar word adding step 18 includes, in the search result of the document full-text search step 14 in which a document is searched using a keyword as a search condition, a document including a synonym / similar word of the search condition keyword in the search result. to add.
That is, a keyword search result including synonyms and similar words as keywords is obtained. At this time, it is also possible to rank the documents of the search result using the similarity between the search condition keyword and the synonym / similar word as the evaluation value of the searched document. When a document containing a synonym / similar word of a synonym / similar word is added to the search result, the product of the similarities can be used as the evaluation value of the searched document.

For example, when performing a full-text search of a document using the keyword “mobile phone”, if “automobile phone” is obtained as a synonym / similar word of “mobile phone” from the synonym / similar word dictionary, A document containing "car phone" in addition to a document containing "phone" is presented to the searcher as a search result. At this time, if the similarity between “mobile phone” and “car phone” is 0.97, “mobile phone”
Assuming that the evaluation value of a document containing one is 1, "car phone"
Can be ranked as 0.97. As described above, even if a document does not include a term that completely matches the keyword input as a search condition of the full-text search, a search result in which documents including a term similar to the input keyword are ranked by similarity of the term Can be presented to the searcher.

Embodiment 8 FIG. FIG. 22 is a flowchart showing the operation of the document search method according to the present embodiment. The same or corresponding parts as in the sixth embodiment are denoted by the same reference numerals and description thereof is omitted.

In the document full-text search similar document adding step 19, if the number of documents belonging to the set of documents of the search result in the document full-text search step 14 is smaller than a predetermined value, the document belonging to the search result document set and the document By adding a document having a high similarity calculated in the similarity calculation step 3 to the search result, the number of documents in the search result is equal to or greater than a predetermined value. For example, it is assumed that the searcher presets to obtain a full-text search result and at least 10 documents and executes a full-text search, but there are only three documents including the keyword input as a search condition. In this case, the document full-text search similar document addition step 19 calculates the similarity between each of the three documents including the keyword input as the search condition and the document not including the keyword input as the search condition by calculating the inner product of the document parameter vector. The document similarity calculation step 3 is performed, and seven documents are added to the search result in order from the document having the highest similarity, thereby presenting 10 documents to the searcher as a search result. As described above, when at least one document is searched by the full-text search, it is possible to present the number of documents desired by the searcher as a search result.

Embodiment 9 FIG. 23 is a flowchart showing the operation of the document search method according to the present embodiment,
Parts that are the same as or correspond to those in Embodiment 6 are given the same reference numerals, and descriptions thereof are omitted. In the figure, a characteristic term attribute determining step 20 evaluates the dependency of the term extracted in the term extracting step 10 on the classification system constructed in the document classifying step 4 to determine a characteristic term having a high appearance frequency in a specific document class. Is extracted.

FIG. 24 is a flowchart for explaining the details of the operation of the characteristic term attribute determining step 20. In steps S2101, S2102, and S2105, all document classes constituting the classification system are processed. S210
3, all documents belonging to a specific document class are processed by S2104 and S2113. S2106, S21
07, S2112 processes all terms of the specific document. The appearance frequency F (W, C) of the term W in the document class C is calculated in S2109. S21
10, the appearance frequency F of the term W in the entire document set
Calculate (W). In S2111, the degree of dependence P (W, C) of the term W on the document class C is calculated. S21
08 to P (W,
Sort in descending order by C). P (W, C) is the degree of dependence of the term W on the document class C, and the higher the term W appears in the document class C and the smaller the term W appears in the entire document, the higher the value.

For example, if 30 documents are registered in a document class of “communication device” and the term “mobile phone” appears 60 times therein, then F (mobile phone, communication device) becomes 60/30 = 2. On the other hand, if the entire document is composed of 1000 documents, and the term “mobile phone” appears 100 times,
(Mobile phone) is 100/1000 = 0.1. Therefore, the degree of dependence P (mobile phone, communication device) of the term “mobile phone” on the document class “communication device” is 2 / 0.1 = 20.
Becomes As described above, in all the document classes, the dependencies of all the terms appearing in the document classes on the document class are calculated, and characteristic terms highly dependent on the document class in each document class can be extracted.

The document full-text search feature term guidance step 21 displays a feature term having a high degree of dependence on the document class obtained in the feature term attribute determination step 20 on the search condition input screen in the document full-text search step 14 on the document condition input screen. Are presented in the form of a term list arranged in descending order, and can be selected and entered as search condition keywords. The above-mentioned characteristic terms are used as keywords, and keyword search based on these characteristic terms can provide the same document search effect as classification search, so even if the searcher has little knowledge of the document classification system, the terms that the searcher knows are used. The search that was used becomes possible. In the classification search 5, characteristic terms may be presented for each document class. As a result, the searcher can grasp the classification system of the document based on information of characteristic terms other than the document class name.

[0078]

As described above, according to the present invention, document parameters are extracted from documents collected from a plurality of sources,
Since the classification management is performed by generating the document parameter vector, there is an effect that the format of the document and the usage of the term are not affected, and the high-quality document classification management and retrieval can be easily performed in a single framework.

Since the appearance ratio of a specific document parameter is indicated at the time of document search using the document parameter as a search condition, there is an effect that the narrowing search of the target document can be easily performed.

Since a plurality of different classification systems are collected and merged and classified and managed, there is an effect that the classification system of the document optimized by the provider can be reflected and a uniform classification system can be constructed. . In addition, there is an effect that documents can be searched over a plurality of classification systems.

Since the classification system can be subdivided, for example, when the number of documents increases, there is an effect that an optimum document search can be performed even if the number of target documents increases.

Since the term extracted from the document is presented as guidance information for each attribute, there is an effect that a document class or a document set to be searched can be searched from the attribute.

Since the synonym / similar word is extracted from the document and the synonym / similar word dictionary is provided, there is an effect that similar documents including the synonym / similar word can be searched.

Since characteristic terms depending on the classification are extracted from the classified documents, a document class or a set of documents to be searched can be searched by the characteristic terms.

[Brief description of the drawings]

FIG. 1 is a flowchart showing an operation of a document classification management and document search method according to Embodiment 1 of the present invention.

FIG. 2 is a diagram illustrating an example of a document to be analyzed and extracted document parameters for explaining the operation of the document classification management method according to the first embodiment;

FIG. 3 is a diagram showing an example of another document and its document parameters, a generated document parameter vector, a classification system, and a document parameter vector.

FIG. 4 is an operation flowchart of document parameter vector generation in document similarity calculation step 3 of FIG. 1;

FIG. 5 is a flowchart showing an operation of a document classification management and document search method according to Embodiment 2 of the present invention.

FIG. 6 is a flowchart showing a detailed operation of a classification parameter search guidance step 7 of FIG. 5;

FIG. 7 is a diagram of a screen showing an example of operations and results for explaining the operation of a classification parameter search step.

FIG. 8 is a flowchart showing an operation of a document classification management and document search method according to Embodiment 3 of the present invention.

FIG. 9 is a flowchart showing a detailed operation of a classification system merging step 8 in FIG. 8;

FIG. 10 is a diagram showing a change in the classification system for explaining the operation of the classification system merging step.

FIG. 11 is a flowchart illustrating an operation of a document classification management and document search method according to Embodiment 4 of the present invention.

FIG. 12 is a flowchart showing a detailed operation of a classification system subdividing step 9 in FIG. 11;

FIG. 13 is a diagram illustrating an example of a relationship between a moving step S and a similarity L between moving documents for explaining the operation of the classification system subdividing step.

FIG. 14 is a flowchart showing an operation of a document search method according to Embodiment 5 of the present invention.

15 is a term attribute determination step 11 in FIG.
FIG. 7 is a flowchart showing a detailed operation of a term attribute dictionary construction step 12.

16 is a diagram illustrating an example of a term extraction rule of FIG. 15 and a result of term extraction applied to a certain document.

FIG. 17 is a diagram illustrating an example of a term attribute extraction result and an example in which the term attribute extraction result is applied to a document search.

FIG. 18 is a flowchart showing an operation of a document search method according to Embodiment 6 of the present invention.

FIG. 19 is a diagram showing an example of an input screen of a full-text search condition in a document full-text search step of FIG. 18;

FIG. 20 is a flowchart showing an operation of a document search method according to Embodiment 7 of the present invention.

21 is a synonym / similar word determination step 16 in FIG.
FIG. 10 is a flowchart showing a detailed operation of synonym / similar word dictionary construction step 17.

FIG. 22 is a flowchart showing an operation of a document search method according to the eighth embodiment of the present invention.

FIG. 23 is a flowchart showing an operation of a document search method according to Embodiment 9 of the present invention.

24 is a flowchart showing a detailed operation of a characteristic term determination step 20 in FIG. 23.

[Description of Signs] 1 document collection step, 2 document parameter extraction step, 3 document similarity calculation step, 4 document classification step, 5 document classification search step, 6 document parameter search step, 7 classification parameter search guidance step,
8 Classification System Merging Step, 9 Classification System Subdivision Step, 10 Term Extraction Step, 11 Term Attribute Determination Step, 12 Term Attribute Dictionary Construction Step, 13 Document Classification Search Guidance Step, 14 Document Full Text Search Step, 15 Document Full Text Search Guidance Step , 16 synonym / similar word determination step, 17 synonym / similar word dictionary construction step, 18 full-text search synonym / similar word addition step, 19 full-text search similar document addition step,
20 characteristic term attribute judgment step, 21 full-text search characteristic term guidance step, 22 classification system acquisition step, 37 common parameter name display, 38 parameter value input area, 39 search target document number, 40 new button, 4
1 Add button, 42 Refine button, 43 Search button, 44 Return button, 45 Distribution button, 46 Parameter selection button, 47 Unique parameter name, 48 Unique parameter containing document number, 49 Input candidate button, 50 Common parameter input value candidate , 51 Common parameter value distribution graph.

 ────────────────────────────────────────────────── ─── Continued on the front page (72) Inventor Yasuhiro Takayama 2-3-2 Marunouchi, Chiyoda-ku, Tokyo Inside Mitsubishi Electric Corporation

Claims (11)

[Claims] 1. A document parameter extracting step of extracting document parameters created by a tag of a predetermined format of a plurality of collected N documents, and the M document parameters extracted from the N documents. For the configured M-dimensional document parameter vector,
Calculating a weight based on the appearance frequency of the component of the parameter vector to determine a document parameter vector for each document; and obtaining a similarity between the document parameter vectors obtained for each document, And a document classification step of classifying the collected documents into a document class of a classification system determined according to degree. 2. A document parameter extracting step of extracting document parameters created by a tag of a predetermined format of a plurality of collected N documents, and M document parameters extracted from the N documents. For the configured M-dimensional document parameter vector,
Calculating a weight based on the appearance frequency of the component of the parameter vector to determine a document parameter vector for each document; and obtaining a similarity between the document parameter vectors obtained for each document, A document classifying step of classifying the documents into a document class of a classification system determined by the degree, and a document extracting a document conforming to the document class indicated in the classification system from a plurality of documents classified and managed in the above-described document classification step. A document search method comprising a classification search step. 3. A document parameter retrieving guidance step for extracting a document parameter having a predetermined appearance ratio close to the specified appearance ratio when an appearance ratio of a specific document parameter to all documents is designated. Document search method described in 2. 4. An average parameter vector of a document class in the same classification system is determined, and if the similarity of the determined average parameter vector between different classification systems is within a predetermined value, the merged classification combining the different classification systems is performed. 2. The document classification management method according to claim 1, further comprising a classification system merging step for setting the same document class in the system. 5. The classification system merging step is characterized in that, when there is a lower-level document class connected to a document class whose similarity is to be obtained, the lower-level connection form is managed as a merged classification system. The document classification management method according to claim 4. 6. A classification system subdividing step for sequentially moving the most similar documents belonging to the same document class and extracting the inter-document similarity for each movement is added. Document classification management method. 7. A term and attribute extracted from a document by a predetermined term extraction rule are constructed and managed as a term attribute dictionary in association with a document class, and when a specific attribute is designated, a term corresponding to the attribute is designated. 3. The document search method according to claim 2, further comprising a document classification search guidance step for extracting a document class having the following. 8. A keyword search mechanism for searching for a document having a specific keyword, and when a specific attribute is designated, the corresponding term is extracted from a term attribute dictionary and searched as a keyword. 7. The document search method according to item 7. 9. The document search method according to claim 7, further comprising a similar document addition step of extracting and adding a predetermined number of documents in descending order of similarity with the search result document class. 10. A synonym / similar word dictionary is constructed and managed based on the similarity between terms of attributes of the same document class, and a keyword search mechanism for searching for a document having a specific keyword is added. 3. The document search method according to claim 2, wherein, in the synonym / similar word dictionary, a term registered and managed equivalently is searched for in addition to the keyword. 11. A feature term determining step of extracting feature terms based on the frequency of appearance in a specific document class, and a keyword search mechanism for searching for a document having a specific keyword are added. 3. The document search method according to claim 2, wherein the characteristic term extracted in the determination step is searched as the keyword.