JP2010218190A - Document retrieval system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a technology capable of improving the accuracy of document retrieval. <P>SOLUTION: A retrieval means retrieves a document similar to a target document and fit for a prescribed condition related to a keyword from a plurality of documents. That is, the retrieval means retrieves the document by the use of both the target document and the keyword. Thus, because both the similar document retrieval and the keyword retrieval are used, prescribed directionality is imparted to a result of the similar document retrieval, so that a probability that the document desired by a user can be obtained is improved. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a technique for searching for a document.

  As a technique related to document search, for example, there are techniques disclosed in Patent Documents 1 and 2.

  According to Patent Document 1, first, any one type of information of a keyword, a document ID, and a document fragment is input, and the input information is sent to a similar document type database. A plurality of feature words included in the document identified from the above are extracted. Then, each extracted characteristic word is sent to the keyword database. Thereby, a document including the characteristic word is searched.

  According to Patent Document 2, first, a target document (search condition document) is input, a plurality of keywords are extracted from the input target document, and a keyword search is performed using each keyword. By the keyword search, a document including the keyword is searched. Then, a document similar to the target document is searched from a plurality of documents found by the search (that is, a similar document search is performed).

JP 2002-222208 A JP 2004-151959 A

  According to Patent Document 1, the information input for the search is any one of a keyword, a document ID, and a document fragment. On the other hand, according to Patent Document 2, the information input for the search is only the target document.

  For this reason, the direction of the search result cannot be specified by any technique described in any of the patent documents. Accordingly, the accuracy of document search is low, and therefore the probability that a document desired by the user is presented is low.

  Accordingly, an object of the present invention is to improve the accuracy of document retrieval.

  Search means searches a plurality of documents for a document that satisfies a predetermined condition relating to a keyword and is similar to the target document. That is, the search means searches for a document using both the keyword and the target document. The keyword may or may not be included in the target document.

  Thus, since the similar document search and the keyword search are used in combination, a certain direction is given to the result of the similar document search, so that it is possible to improve the probability of obtaining the document desired by the user.

  Note that “a document that satisfies a predetermined condition regarding a keyword” is, for example, a document that includes or does not include k keywords (k is a natural number). Typically, k is considered to be 1, but k may be 2 or more. In this case, for example, if the condition is “includes two keywords“ ABC ””, a document including only one keyword “ABC” does not correspond to “a document that satisfies a predetermined condition regarding keywords”. It is.

  The “document” may be any data as long as it is electronic data including characters (for example, it may be a document including a table or an image).

FIG. 1 shows a computer system having a mail audit system to which a document search system according to a first embodiment of the present invention is applied. FIG. 2 shows an example of the flow of search processing in the first embodiment of the present invention. FIG. 3 is an explanatory diagram of the secondary search in the search process according to the first embodiment of the present invention. FIG. 4 shows a computer system having a mail auditing system to which the document search system according to the second embodiment of the present invention is applied. FIG. 5 is an explanatory diagram of search processing in the second embodiment of the present invention. FIG. 6A is an explanatory diagram of a first example of a method for selecting at least one mail from a mail group classified into a similar category in a similar mail space (hereinafter referred to as a mail selection method). FIG. 6B is an explanatory diagram of a second example of the mail selection method. FIG. 6C is an explanatory diagram of a third example of the mail selection method. FIG. 6D is an explanatory diagram of a fourth example of the mail selection method. FIG. 7 shows an example of the flow of search processing in the second embodiment of the present invention. FIG. 8 is a supplementary diagram for explaining the mail selection method shown in FIG.

  Hereinafter, taking a case where a document is an electronic mail (hereinafter simply referred to as “mail”) as an example, some embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 1 shows a computer system having a mail audit system 103 to which a document search system according to the first embodiment of the present invention is applied.

  Mail is transmitted and received between the internal terminal 101 and the external terminal 105 via the internal network 111 and the external network 112. The in-house terminal 101 and the outside terminal 105 are, for example, personal computers or server machines. The in-house network 111 is, for example, a LAN (Local Area Network). The external network 112 is, for example, a network including an external intranet and the Internet different from the internal network 111.

  A mail server 107 and a mail audit system 103 are connected to the in-house network 111.

  The mail server 107 stores e-mails transmitted / received via the in-house network 111.

  The mail auditing system 103 checks e-mails that pass through the in-house network 11, in particular, e.g., e-mails that are transmitted from the company to the outside (so-called outbound mails).

  The mail audit system 103 includes, for example, a CPU 131, storage resources (for example, the memory 132 and the storage device 135), and an interface device (communication I / F) 133 that controls communication with an external device.

  The memory 132 stores, for example, various computer programs, various data used for processing performed by the CPU 131, and the like. An example of the computer program is a search program 321.

  The CPU 131 can perform a mail search by executing a computer program stored in the memory 132. Specifically, for example, the CPU 131 implements the function of the search unit 301 by executing the search program 321.

  The storage device 135 stores a similar mail database (similar mail DB) 441 and a keyword database (keyword DB) 442. The similar mail DB 441 is a DB used for similar mail search. The keyword DB 442 is a DB used for keyword search.

  In the present embodiment, the search unit 301 searches for mail from the mail server 107 using both the keyword and the target mail. The search unit 301 receives, for example, from a user via a user terminal that can communicate with the mail audit system 103 (for example, the in-house terminal 101 or a communication terminal that is connected to the mail audit system 103 without going through the in-house network 111). A search instruction can be received, a search process can be performed in response to the instruction, and the search result can be output to the user terminal. The output search result is displayed on the display screen of the user terminal.

  FIG. 2 shows the flow of search processing performed by the search unit 301.

  A keyword is input (S101). This keyword is, for example, a keyword input by the user. Of course, the keyword is not limited thereto, and may be a keyword arbitrarily selected from a keyword list stored in the memory 132, for example.

  The search unit 301 performs a primary search (keyword search) using the input keyword (S102). Specifically, for example, the search unit 301 searches for a mail including the input keyword from a plurality of mails included in the mail server 107. As a result, for example, it is assumed that ten mails (group A) are retrieved from 10,000 mails (group C) that the mail server 107 has.

  At least one mail is automatically selected from the 10 mails (group A) found by the search in S102 (S103). As a selection criterion, for example, there is a criterion for how many keywords input in S101 are included. Here, it is assumed that the mail X including the most keyword input in S101 and the mail Y including the second most keyword are selected.

  The mail selected here is the target mail. The “target mail” is a mail that is a target (starting point) of similar mail search.

  The search unit 301 performs a secondary search that is both a keyword search using a keyword and a similar email search using a target email. In this secondary search, a mail that does not include the keyword input in S101 and is similar to the target mail is searched. As a result, for example, it is assumed that 18 mails similar to the target mail X and 22 mails similar to the target mail Y are retrieved from the group B (9990 mails) other than the group A in the group C.

  The search unit 301 outputs the search result (S105). This search result includes, for example, the number of search hits and a mail list. The number of search hits is, for example, the total of the number of emails found in the primary search and the number of emails found in the secondary search. The mail list includes, for example, information related to mails found in the primary search (for example, sender and destination mail addresses and mail text) and information related to mails found in the secondary search.

  The above is the flow of search processing. This flow is an example, and for example, any of the following modifications may be adopted. That is, in S102, a mail that does not include the keyword input in S101 may be searched. In S103, the user may select the target mail from group A. In S104, a keyword different from the keyword input in S101 (for example, a keyword included in the target mail selected in S103 or a keyword input separately by the user) may be used. In S104, an email including a keyword may be searched.

  The secondary search will be described in detail with reference to FIG.

  In the secondary search, the similar mail DB 441 and the keyword DB 442 are used.

  The similar mail DB 441 is a table representing the relationship between each similarity and the mail. The similar mail search in the present embodiment is a similar mail search using LSH (Locality Sensitive Hashing), that is, an approximate neighborhood search method. For this reason, the mail similarity is a mail hash value.

  The keyword DB 442 is a table that represents the relationship between each keyword and mail that includes the keyword.

  In the present embodiment, each mail 451 in the existing mail group (mail group stored in the mail server 107) 443 is registered in advance (for example, in a night batch) in the similar mail DB 441 and the keyword DB 442. Specifically, for the similar mail DB 441, the hash values of all the mails 451 are calculated, and the ID of the mail 451 is added to the column corresponding to the hash value (the hash value is similar mail). If it is not registered in the DB 441, its hash value and mail ID are registered in the similar mail DB 441). On the other hand, regarding the keyword DB 442, a word is extracted from the mail 451, and the ID of the mail 451 is added to the column corresponding to the same keyword as the word (if the extracted word is not registered in the keyword DB 442). The word (keyword) and the mail ID are registered in the keyword DB 442).

  In the secondary search, similar mail search, keyword search, and search result integration are performed. In the similar mail search, the search unit 301 calculates a hash value of the target mail, and acquires all mail IDs (first mail ID group) corresponding to the hash value from the similar mail DB 441. In the keyword search, the search unit 301 acquires all mail IDs (second mail ID group) corresponding to the keywords from the keyword DB 442. In the search result integration, for example, if the condition is “does not include the specified keyword”, the search unit 301 sends a mail that is different from the mail ID included in the second mail ID group from the first mail ID group. All IDs are acquired (if the condition is “include specified keyword”, the search unit 301 uses the same mail ID as the mail ID included in the second mail ID group from the first mail ID group. All). The acquired mail ID is the ID of the mail hit in the secondary search. If the target mail is the target mail X in FIG. 2, the ID of the mail hit in the secondary search is the ID of the 18 mails.

  The above is the description of the first embodiment. The similar mail search is not limited to the similar mail search according to the LSH technique, and a similar mail search according to another technique may be employed.

  According to the first embodiment described above, since the similar mail search and the keyword search are used in the secondary search, a certain directionality is given to the result of the similar mail search. Therefore, it is possible to improve the probability of obtaining the mail desired by the user. Specifically, for example, in the following first and second cases, the probability of obtaining a user's desired mail is improved.

First case: Prevent search omission in keyword type audit.
“Keyword-type audit” refers to, for example, extracting outgoing mail containing NG keywords registered in a prepared NG keyword list, and manually checking the text of the extracted outgoing mail This is a method for auditing whether outgoing e-mails (hereinafter referred to as dangerous e-mails) that are disadvantageous to the company have leaked outside the company. The present embodiment can be applied to this case as follows. That is, the search unit 301 searches a mail including the NG keyword “XXX” by a primary search, sets a mail (dangerous mail) selected from the mail group found by the primary search as a target mail, and “similar to the target mail”. A secondary search is performed for a mail that does not include the keyword “ABC”. Thereby, it is possible to find a mail similar to the dangerous mail including the NG keyword “ABC” from the mail group not including the NG keyword “ABC”. That is, the completeness of the search is ensured and the certainty of detection of dangerous mail outflow is improved.

Second case: The search result in the keyword search is narrowed down.
The present embodiment can be applied to this case as follows. That is, the search unit 301 searches a mail including the keyword “DFG” from the user by a primary search, and selects a mail selected from the mail group found by the primary search (user desired mail) as a target mail. A secondary search is performed for a mail similar to the mail and including the keyword “DFG”. In other words, in addition to the keyword “DFG” input by the user, the similar mail search is used for narrowing down. As a result, the search result can be narrowed down so that the number of mails other than the user-desired mails is reduced without the user newly inputting a keyword.

  A second embodiment of the present invention will be described. At that time, differences from the first embodiment will be mainly described, and description of common points with the first embodiment will be omitted or simplified. In the following description, a mail with mail ID = x (x is an integer) is represented as “mail #x”.

  FIG. 4 shows a computer system having a mail audit system 803 to which the document search system according to the second embodiment of the present invention is applied.

  A plurality of similar mail DBs 441 are prepared in the storage device 135. In other words, in the present embodiment, a plurality of similar mail spaces are defined. In the present embodiment, it is assumed that, for example, similar mail spaces A and B are defined as a plurality of similar mail spaces. For this reason, it is assumed that similar mail DBs 441A and 441B are prepared as the similar mail DB 441.

  The search unit 801 (search program 821 executed by the CPU 131) realized by the CPU 131 has a function of performing a similar mail search that follows the similar mail space instead of or in addition to the function of the search unit 301.

  FIG. 5 is an explanatory diagram of similar mail search performed by the search unit 801.

  Similar mail spaces A and B are spaces defined based on different similarity models of LSH. That is, each hash value (category) in the similar mail space A (first similar mail DB 441A) and each hash value (category) in the similar mail space B (second similar mail DB 441B) are calculated according to different similarity models. Has been. For example, the hash value for the similar mail space A is obtained by a method according to the similarity model A, and the hash value for the similar mail space B is obtained by a method according to the similarity model B.

  For each mail in the mail group (existing mail group) stored in the mail server 107, each hash value is calculated in advance according to each similarity model of the similar mail spaces A and B. Each mail is classified into the similar mail spaces A and B according to the respective hash values.

For example, when a target mail is input, the search unit 801 can search mail # 4 as a mail similar to the target mail by performing the following processes (5-1) to (5-6). it can.
(5-1) The hash value = 48 of the target mail is calculated according to the similarity model A of the similar mail space A that is the initial search range of the target mail.
(5-2): A mail group corresponding to hash value = 48 is searched from the similar mail space A.
(5-3): Mail # 5 is selected by a predetermined method from the mail group found in the search of (5-2).
(5-4): Using ID = 5 of the mail # 5 selected in (5-3) as a key, a similar mail space B which is a different search range from the search range immediately before (5-4) refer. As a result, a mail group having a hash value = 948 in which the mail # 5 is classified is found from the similar mail space B.
(5-5): Mail # 8 is selected from the mail group found in (5-4) by a predetermined method.
(5-6): Using ID = 8 of the mail # 8 selected in (5-5) as a key, a similar mail space A that is a different search range from the search range immediately before (5-6) refer. As a result, a mail group of hash value = 18 in which the mail # 8 is classified is found from the similar mail space A.
(5-7): Mail # 4 included in the mail group found in (5-6) has a hash value 483 according to the similarity model B. In the similar mail space B that is the search range immediately before (5-6), mail # 1 is also classified with hash value = 483. Mail # 1 is mail classified into the same hash value = 48 as the target mail in the similar mail space A in the first search range. From the above, the mail # 4 is similar to the mail # 1 similar to the target mail, and the mail # 1 is similar to the target mail. For this reason, mail # 4 is determined to be mail similar to the target mail.

  As described above, it is possible to find mail # 4 that is dissimilar to the target mail in the similar mail space A in the first search range, but similar to the target mail in another similar mail space B.

  A mail similar to the target mail is not found even if only one similar mail space A is referred to. In other words, a mail found by referring to only one similar mail space A based on the target mail is not necessarily similar to the target mail.

  In this embodiment, a plurality of similar mail spaces according to a plurality of similarity models are defined, each mail is classified into each similar mail space, and the similar mail space is traced based on the target mail (in the above example, similar mail spaces are traced). The mail spaces A and B are alternately referred to). That is, a mail similar to the target mail is searched from a plurality of viewpoints. Thus, it is possible to search for mail # 4 that is actually similar to the target mail. In other words, the accuracy of similar mail search can be improved as a whole.

  The above processing is an example, and for example, any one of the following modifications may be adopted.

  For example, in the above example, the similar mail space A is referred to first, but which similar mail space is set as the first reference destination may be defined in advance, or may be changed randomly or by another method. May be.

  Further, for example, when searching for a mail group classified into the corresponding hash value (for example, when searching for (5-2) or (5-4) above), as in the first embodiment Keyword search may be used in combination. Specifically, for example, in the search of (5-1), mails that are classified as hash value = 48 and do not include (or include) the keyword “HIJ” are searched. The keyword “HIJ” may be a keyword extracted from the target mail, or may be a keyword that is not included in the target mail and input from the user. Further, for example, in the search of (5-3), mails that are classified as hash value = 948 and do not include (or include) the keyword “KLM” are searched. The keyword “KLM” is either a target mail or a mail that has been searched (or selected) so far (for example, any mail classified as hash value = 48 or selected mail # 5). The keyword may be extracted from the above, or may be a keyword input by the user separately.

  In addition, for example, keyword search may be adopted in the final narrowing down. That is, a mail that does not include (or includes) the keyword “NOP” may be searched from a mail group similar to the target mail acquired as a result of tracing the similar mail space based on the target mail. The keyword “NOP” is either a target mail or a mail that has been searched (or selected) so far (for example, any mail classified into the hash value of space A = 48, the hash value of space B = It may be a keyword extracted from any mail classified as 483, selected mail # 5 or mail # 8), or may be a keyword input by a user separately.

  Further, for example, when tracing a similar mail space, keyword search may not be used at all. In other words, the keyword DB 442 may not be provided.

  The plurality of similar mail spaces may include similar mail spaces according to different types of similar mail search methods. In the above example, in addition to a similar mail space according to LSH, a similar mail space according to another similar mail search method (for example, mail such as “business”, “private”, etc., not a hash value as a category) The space in which the type is adopted may be included.

  By the way, the “predetermined method” in (5-3) and (5-5) (that is, the method of selecting one mail # 5 from the hash value = 48 of the similar mail space A and the hash of the similar mail space B) The method of selecting one mail # 8 from value = 948) is, for example, one of the methods (A) to (D) shown in FIG.

  The method of FIG. 6A is a method of selecting by a keyword. According to the example of FIG. 6A, mail # 5 that does not include (or includes) the input keyword is selected from the mail group classified as hash value = 48 in space A. The input keyword may be a keyword extracted from either the target mail or the mail searched (or selected) so far, or may be a keyword input by a user separately. .

  The method of FIG. 6B is a method in which the user manually selects. According to the example of FIG. 6B, the user browses each mail classified as hash value = 48 (for example, browses the body of each mail) and selects a desired mail # 5.

  The method of FIG. 6C is a method of selecting by full-text word search. According to the example of FIG. 6C, various words are extracted from all mail classified as hash value = 48, word statistics for each mail and word statistics for the entire mail group are calculated, and the mail group Mail # 5 is selected based on the entire word statistics and the word statistics for each mail. Specifically, for example, according to the word statistics of the entire mail group, the mail having the largest number of words “replacement” and the largest number of words “replacement” is the mail group classified as hash value = 48. Mail # 5 is selected for Mail # 5. In the full text word search, words included in the target mail may be taken into consideration.

  The method of FIG. 6D adjusts the distance of the radius R around the input mail so that the number of acquired mails is P (P is a natural number) and is within the range of the radius R. This is a method of selecting mail. The “input mail” is the target mail in (5-2), and is mail # 5 in (5-4). This input mail is converted into a vector as shown in FIG. 8 and projected to a similar mail space. The conversion into a vector may be performed in any way. For example, here, a word included in a mail sentence is used. The words included in the mail text are compared with a word list separately listed in advance, and “1” is included for those included in the word list, and “0” is included for those not included. This forms a vector. Note that different similar mail spaces are formed by changing the types of words to be listed. According to the example of FIG. 6D, by reducing the distance of the radius R around the target mail W in the similar vector space from R1 to R2, the mail group classified as hash value = 48 is reduced to one. Narrowing down to mail # 5 is performed.

  In addition to words, elements used for vector conversion may be email transmission time, attached file, or transmission format (new email, forwarding, reply, etc.). . Even when using words, the number of included words may be used instead of the presence or absence of words.

  At least one of the methods of FIGS. 6A to 6D described above can narrow down the search results of the mail group classified into the corresponding hash value (for example, the above (5-2) and ( (5-4) Search result refinement). For example, in a case where 10,000 mails are classified into hash value = 48 in the similar mail space A at the beginning of the target mail, in order to narrow down 10,000 mails, FIG. 6 (A) to FIG. 6 (D). At least one of the methods may be used.

  FIG. 7 shows the flow of search processing performed by the search unit 801. In the following description, for the sake of easy understanding, the hash value and the mail ID shown in FIG. 5 are used as appropriate.

  When the target mail is input (S701), the search unit 801 calculates a hash value (for example, 48) of the target mail according to the similarity model of the similar mail space (space A) that is the first search range of the target mail. (S702).

  Next, the search unit 801 searches the first similar mail space (space A) for a mail group classified into the same hash value as the calculated hash value (S703).

  Next, the search unit 801 selects a mail (mail # 5) selected from the mail group found by the previous search from a similar mail space (space B) different from the similar mail space (space A) in the previous search range. The mail group classified by the hash value to which () belongs is searched (S704). Specifically, for example, the search unit 801 refers to another similar mail space by using the mail ID of the selected mail as a key, so that the hash to which the selected mail belongs from the other similar mail space. Search mail group classified by value. The “selected mail” is a mail selected by any one of the methods shown in FIGS. 6A to 6D from the mail group found by the previous search (this is the following) The same applies to the explanation).

  Next, the search unit 801 determines whether or not the mail (mail # 8) selected from the mail group found by the search in S704 has a predetermined relationship with the target mail (S705). In other words, the search unit 801 determines whether or not the similar mail search can be terminated.

  If the result of the determination in S705 is negative (S705: NO), the search unit 801 re-executes S704. Specifically, for example, since the search range of the immediately previous S704 was the similar mail space B, the search unit 801 uses another similar mail space A as the key with ID = 8 of the mail selected immediately before. refer. As a result, a mail group classified as hash value = 18 to which mail # 8 belongs can be found from another similar mail space A.

  As described above, the search unit 801 repeats S704 until a positive result is obtained as a result of the determination in S705. Therefore, the search range immediately before in S704 is the search range in S703 (similar mail space A that is the first search range) or the search range in the previous S704.

  If the determination result in S704 is affirmative (S704: YES), the search unit 801 determines that the mail (mail # 4) found by the last search in S704 is similar to the target mail (S706). .

  Here, the “predetermined relationship” in the determination in S705 is the hash value (= 48) in the mail (mail # 4) found by the search in S704 and the initial search range (similar mail space A) of the target mail. That is, the mail (mail # 1) classified into the same hash value (= 483) in the search range (similar mail space B) immediately before S704.

  With such a search process, the search described with reference to FIG. 5 is performed. That is, the similar mail spaces A and B are alternately referred to, and as a result, the mail # 4 that is actually similar to the target mail is searched.

  As described above, according to the second embodiment, by combining a plurality of similar mail searches, it is possible to compensate for the low accuracy of individual similar mail searches and to improve the accuracy of similar similar mail searches as a whole. it can.

  The several embodiments of the present invention described above are examples for explaining the present invention, and are not intended to limit the scope of the present invention only to those embodiments. The present invention can be implemented in various other modes without departing from the gist thereof.

103,803 ... Mail audit system

Claims (9)

A keyword input means for inputting keywords;
A target document input means for inputting the target document;
A document search system comprising: a search unit that searches a plurality of documents for a document that satisfies a predetermined condition relating to the keyword and that is similar to the target document. The keyword input means receives a keyword from a user,
The search means performs a primary search for searching N documents (N is a natural number) including the keyword from a plurality of documents;
After the primary search, the target document input means selects and inputs M target documents (M is a natural number and less than or equal to N) from the N documents, and the search means inputs the target document. A secondary search for searching L documents (L is a natural number) that are similar to and not including the keyword from the plurality of documents,
The search means outputs a search result including information representing the N documents and the L documents;
The document search system according to claim 1. The keyword input means inputs a first keyword, and the search means performs a primary search for searching N documents (N is a natural number) including the first keyword from a plurality of documents.
The keyword input means inputs a second keyword, and the target document input means inputs M target documents (M is a natural number and less than or equal to N) selected from the N documents, The search means performs a secondary search that searches the plurality of documents for a document that meets a predetermined condition related to the second keyword and is similar to the target document;
The document search system according to claim 1. The predetermined condition regarding the second keyword is a condition that the second keyword is not included.
The document search system according to claim 3. There are different document spaces that follow different perspectives, and each document space has multiple similar categories that are determined based on that document space perspective,
Each document is classified into one of the similar categories in any two or more document spaces,
The search means performs the following processes (A) to (D):
(A) identifying a similar category of the target document in the document space based on the viewpoint of the document space that is the first search range of the target document;
(B) searching the first document space for documents classified in the same similar category as the identified similar category;
(C) Search a document classified in the same similar category as the document found by the process immediately before (C) from a document space different from the search range immediately before the process of (C);
(D) determining whether the document found by the process of (C) has a predetermined relationship with the target document;
Run
If the result of the determination in (D) is negative, the search means re-executes the process in (C),
If the result of the determination in (D) is affirmative, the search means determines that the document found by the process in (C) is a document similar to the target document;
The document found by the processing of (B) and / or (C) is a document that meets a predetermined condition regarding keywords.
The document search system according to any one of claims 1 to 4. The predetermined relationship is that a document found by the process (C) and a document classified into a similar category in the first document space of the target document are searched immediately before the process (C). Being in the same similar category in the document space of the scope,
The document search system according to claim 5. The document found from the search range immediately before the process (C) is a document in which two or more documents found from the search range are narrowed down using keywords.
The document search system according to claim 5 or 6. At least one document space is a space based on a similarity model of LSH (Locality Sensitive Hashing),
Each similarity category is a hash value,
The document found from the search range immediately before the process of (C) is the input of a plurality of documents having the same hash value as the document input for the search in (C). A document belonging to a range within an adjusted radius R centered on the document;
The document search system according to claim 5 or 6. Entering keywords,
Entering the target document; and
A computer program for causing a computer to execute a step of searching a plurality of documents for a document that satisfies a predetermined condition related to the keyword and is similar to the target document.

Images