JP4230710B2 - SEARCH DEVICE, SEARCH METHOD, AND PROGRAM - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a search device suitable for performing a high-speed search, a search method, and a program for causing a computer to implement these search methods.
[0002]
[Prior art]
2. Description of the Related Art Conventionally, a search system has been proposed in which a search target document is registered in a database, and when a search request is given to the database, a document that conforms to the search request is output as a search result.
[0003]
Even in the field of Internet technology, which has been remarkably popular in recent years, the URL (Universal Resource Locator) of websites that collect information on websites around the world, create a database, and adapt to search requests from users. Internet search engines are also provided by each company.
[0004]
[Problems to be solved by the invention]
However, in order to enable a large-scale search, it is necessary to increase the size and number of databases. In such a case, the input / output and communication amount with the database, and the processing amount of the database itself increase. As a result, the time (waiting time) required for the search becomes long and the load on each database system may increase.
[0005]
Japanese Patent Application Laid-Open No. 11-39324 discloses a search technique in which a plurality of search subsystems are prepared, a tendency of a search request issued by a user is analyzed in advance, and one of the search subsystems is selected to perform load distribution. Although disclosed, depending on the situation, the effect of this technique may not be obtained. Therefore, there is a great demand for various load distribution techniques to cope with various situations.
[0006]
The present invention solves the above-described problems, and an object of the present invention is to provide a search device and a search method suitable for performing high-speed search, and a program for causing a computer to implement them.
[0007]
[Means for Solving the Problems]
In order to achieve the above object, the following invention is disclosed in accordance with the principle of the present invention.
[0008]
A search device according to a first aspect of the present invention includes a group division unit, an index unit, an input reception unit, a group score calculation unit, a group selection unit, a search unit, and a merge unit, The configuration is as follows.
[0009]
That is, the group dividing unit divides the search target document into a plurality of groups.
[0010]
On the other hand, the index unit, for each of the plurality of divided groups, information of the group, the keyword that appears in the search target document included in the group, and the number of search target documents in which the keyword appears Store in association with each other.
[0011]
Further, the input receiving unit receives an input of a search request.
[0012]
And a group score calculation part calculates the group score with respect to the search request | requirement which was input with reference to the information memorize | stored in the index part about each of the some divided | segmented group.
[0013]
On the other hand, the group selection unit selects one or more groups from a plurality of groups based on the calculated group score.
[0014]
Further, the search unit searches in parallel for each of the selected one or more groups in response to a search request that has been accepted as an input.
[0015]
Then, the merging unit merges the results searched for each of the selected one or more groups, and outputs this as a search result for the search request for which the input has been accepted.
[0016]
According to the present invention, the search target documents are classified into groups based on a predetermined criterion in advance, a group score corresponding to the criterion is calculated from the search request, and a group to be searched is selected thereby, and load distribution is achieved. The user can obtain search results with a short waiting time.
[0017]
In the search device of the present invention, the index unit further stores, for each of the plurality of divided groups, information on the group and the total number of words included in the search target document included in the group in association with each other. It can be constituted as follows.
[0018]
Further, in the search device of the present invention, the index unit, for each of the plurality of divided groups, for the group and all keywords included in the search target document included in the group, the keyword, the group, and the group The number of documents in which the keyword appears can be stored in association with each other.
[0019]
These inventions are one of the preferred embodiments of the above invention, and the calculation of the group score includes “the total number of words included in the search target document included in the group”, “the number of appearance documents of the keyword included in the group”, etc. Information can be used.
[0020]
The search device of the present invention can be configured as follows.
[0021]
That is, for each of one or more selected groups, the search unit calculates a document score according to a search request that is accepted for input to a document included in the group.
[0022]
On the other hand, the merging unit refers to the calculated group score and the calculated document score, calculates the total score, and outputs the search results for the search request accepted for input in the order of the total score. .
[0023]
According to the present invention, it is possible to appropriately determine the order in which the search results are presented to the user using the total score, and to present the search results in the order according to the desire desired by the user.
[0024]
In the search device of the present invention, the merging unit merges the results searched for each of the selected one or more groups in the order in which the search target documents included in the results are obtained, and outputs them in that order. It can be constituted as follows.
[0025]
According to the present invention, when retrieval processing is performed in a plurality of database systems, it is possible to reduce the waiting time by presenting the retrieval target document to the user in the order in which the retrieval target documents are obtained.
[0026]
The search device of the present invention can be configured as follows.
[0027]
That is, each search target document is associated with a URL.
[0028]
On the other hand, when the number of search target documents included in each group exceeds the number of desired ranges, the group division unit divides the URL tree structure associated with the search target documents in the width direction. Subdivide the group.
[0029]
Further, when the number is less than the desired number, the group and another group are merged so that the number of search target documents included in the group after merging is equal to or less than the number of desired ranges.
[0030]
According to the present invention, an appropriate number of groups can be configured based on the URL tree structure.
[0031]
Further, in the search device of the present invention, the group dividing unit calculates a cosine of a vector composed of the frequencies of a plurality of words appearing in each search target document and a predetermined vector, and depending on the calculation result The search target document can be configured to be classified.
[0032]
According to the present invention, a feature of a document is expressed as a vector, and a cosine having information corresponding to an angle between the vector and a predetermined vector is calculated to calculate a “similarity” (how much) to the predetermined vector. The document can be divided into groups depending on whether the corner is small.
[0033]
In the search device of the present invention, the group score calculation unit performs a predetermined similarity calculation (including calculation of the cosine of both) on the search request that has been accepted and the predetermined vector. This can be configured as a group score.
[0034]
According to the present invention, the feature of the search request is represented by a vector, and the search request is classified into one or more groups based on the “similarity” between this and the “predetermined vector” used for document classification. It becomes possible to sort.
[0035]
In the search device of the present invention, the group dividing unit is configured to calculate a document score for each search target document according to a predetermined search request and classify the search target document according to the calculation result. be able to.
[0036]
According to the present invention, the document score used to determine the order in which the documents are arranged can be used for grouping.
[0037]
A search method according to another aspect of the present invention includes a group division process, an index process, an input reception process, a group score calculation process, a group selection process, a search process, and a merge process. Configure as follows.
[0038]
That is, in the group dividing step, the search target document is divided into a plurality of groups.
[0039]
On the other hand, in the indexing process, for each of the plurality of divided groups, information on the group, the keyword appearing in the search target document included in the group, and the number of search target documents in which the keyword appears is associated. Add and remember.
[0040]
Further, in the input receiving process, an input of a search request is received.
[0041]
In the group score calculation step, for each of the plurality of divided groups, a group score for a search request that is accepted by referring to stored information is calculated.
[0042]
On the other hand, in the group selection step, one or more groups are selected from a plurality of groups based on the calculated group score.
[0043]
Further, in the search step, search is performed in parallel for each of the selected one or more groups in response to a search request that has been accepted as input.
[0044]
In the merging step, the search results for each of the selected one or more groups are merged and output as search results for the search request that has been accepted as input.
[0045]
A program according to another aspect of the present invention is configured to cause a computer to function as the search device described above or to cause the computer to execute the registration method described above.
[0046]
The above program can be recorded on various recording media such as CD-ROM (Compact Disk Read Only Memory) and FD (Flexible Disk) and distributed via a computer communication network such as the Internet. Can do.
[0047]
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of the present invention will be described below with reference to the accompanying drawings.
[0048]
(Outline configuration of information processing device)
FIG. 1 is a schematic diagram showing a schematic configuration of a typical information processing apparatus that can function as a sub apparatus constituting the search apparatus of the present invention by processing a recording medium in which the data structure of the present invention is recorded. Hereinafter, a description will be given with reference to FIG.
[0049]
The information processing apparatus 101 is controlled by a CPU (Central Processing Unit) 102. When the information processing apparatus 101 is turned on, the CPU 102 executes an IPL (Initial Program Loader) stored in the ROM 103.
[0050]
The IPL reads and executes an OS (Operating System) program stored in a recording medium such as the hard disk 104, the FD attached to the FD drive 110, and the CD-ROM attached to the CD-ROM drive 111. It is a program.
[0051]
After starting the OS, the CPU 102 executes an application program stored in the hard disk or the like in accordance with a user instruction input with the keyboard 105 or the mouse 106 or in accordance with the contents of a setting file described in advance on the hard disk or the like. Execute.
[0052]
By executing the application program, the information processing apparatus 101 functions as a sub apparatus.
[0053]
As a result, the management information of the sub device can be input, and the sub device responsible for the dialogue with the user can accept the input of the search request from the user.
[0054]
The CPU 102 uses the RAM 107 as a temporary work storage area when executing these programs. In addition, a register or a cache (not shown) provided in the CPU 102 is used as a temporary working storage area.
[0055]
As the program is executed, the CPU 102 reports the result to the user, shows the progress of the program, and presents management information. Can be displayed. In the instruction operation using the mouse 106, the cursor displayed on the screen is moved by moving the mouse 106, and the menu item pointed to by the cursor can be selected by clicking the mouse 106.
[0056]
The information processing apparatus 101 can communicate with each other via an interface 109 such as a NIC (Network Interface Card) or a modem, and can communicate with a computer communication network such as the Internet. In addition, a program received via the interface 109 can be executed.
[0057]
(Embodiment of the present invention)
In the following description, for ease of understanding, an embodiment of the present invention will be described by taking as an example a case where the present invention is applied to an Internet search engine. FIG. 2 is a schematic diagram showing a schematic configuration of the search device according to the present embodiment. Hereinafter, a description will be given with reference to FIG.
[0058]
The search device 201 of the present embodiment includes a group dividing unit 202, an index unit 203, an input receiving unit 204, a group score calculating unit 205, a group selecting unit 206, a searching unit 207, and a merging unit 208. Prepare.
[0059]
FIG. 3 is a flowchart showing the flow of index creation processing executed in the search device according to the present embodiment.
[0060]
The group dividing unit 202 divides the search target document into a plurality of groups (step S301). In this embodiment, the search target document is an HTML (Hyper Text Markup Language) document that exists on each website in the Internet. This is divided into a plurality of groups by the following method.
[0061]
The group division unit 202 calculates, for each of the search target documents, a cosine between a vector including a frequency of a plurality of words appearing therein and a predetermined vector, and classifies the search target document according to the calculation result. It is.
[0062]
Consider m vectors that are expected to be frequently used for search in advance, and consider a vector whose frequency is the frequency at which each keyword appears in the search target document.
[0063]
For example, when dividing into n groups, n unit vectors different from each other are appropriately prepared as “predetermined vectors”.
[0064]
Then, by calculating the cosine of each unit vector and the vector of the search target document, the one with the smallest angle between the unit vector and the vector of the search target document can be selected. Then, the search target document may be classified into a group associated with the selected unit vector.
[0065]
Other classification methods will be described later as other embodiments.
[0066]
On the other hand, for each of the plurality of divided groups, the index unit 203 includes information on the group, the keyword that appears in the search target document included in the group, and the number of search target documents in which the keyword appears. The data is stored in association with each other (step S302). As this keyword, the same keyword used in the above-mentioned group classification can be adopted.
[0067]
The index unit 203 can calculate a group score at high speed by using a transposed index table of various pieces of information of necessary documents. In the inverted index table, when a keyword (or an ID corresponding to the keyword) included in the search request is given, a position where the keyword appears and statistical information can be obtained. The inverted index table can be constructed by a known technique such as hash, B-tree, or trie structure using keywords as keys.
[0068]
The search unit 207 is provided with a database (DB) associated with each group, and the search target document itself belonging to each group together with the URL of the search target document and the like is also included in this database. Or a pre-processed version of this.
[0069]
Information that is desirably stored other than this will be described later as other embodiments.
[0070]
FIG. 4 is a flowchart showing the flow of search response processing executed in the search device according to the present embodiment.
[0071]
The input reception unit 204 receives a search request input (step S401). The search request to be accepted is preferably normalized so as to be expressed by an appropriate keyword. Moreover, you may weight a keyword suitably. For normalization, various database search techniques can be used.
[0072]
Then, the group score calculation unit 205 calculates, for each of the plurality of divided groups, a group score for a search request that has been accepted by referring to information stored in the index unit 203 (step S402).
[0073]
In this embodiment, the cosine of the vector is also used for calculating the group score. That is, for each of the n groups, a unit vector in the same direction as a vector whose element is the number of search target documents in which the keyword appears is considered. On the other hand, for a search request, a vector whose element is the number of keywords appearing in the search request (weighted) is obtained. Then, the cosine of each unit vector and the obtained search request vector is calculated, and this is used as the group score. In the case of this method, the value of the group score is 0 or more and 1 or less, and the larger the value, the smaller the angle between the two vectors.
[0074]
On the other hand, the group selection unit 206 selects one or more groups from the plurality of groups based on the calculated group score (step S403).
[0075]
In the present embodiment, a predetermined number of groups are selected from those having a large group score.
[0076]
Further, the search unit 207 searches in parallel for each of the selected one or more groups in response to a search request that has been accepted (step S404).
[0077]
A DB is assigned to each group in the search unit 207, and a system (Q) for performing a dedicated search is assigned to each group. Therefore, each Q can execute a search for each DB in parallel.
[0078]
The group to which the search target document that the user wants belongs is very likely to have a small angle between two vectors, so if database search processing is performed only on these, the necessary information can be almost obtained. It becomes. On the other hand, since unnecessary database search processing is not performed, load distribution can be appropriately achieved. In addition, since the database search process is performed in parallel, a search with a short waiting time is possible.
[0079]
Then, the merging unit 208 merges the results searched for each of the selected one or more groups, and outputs this as a search result for the search request that has been accepted (step S405).
[0080]
The following can be considered as a method of merging.
[0081]
Most simply, the search results for each of the selected one or more groups are merged in the order in which the search target documents included in the results are obtained, and output in that order.
[0082]
One or more of the above-mentioned sub devices may be assigned to each of these units, or one or more of the above-mentioned sub devices may be assigned to any combination of these units. However, it is desirable to assign the same number of sub-devices as the number of groups to the search unit 207 so that searches in each group can be independently performed in parallel.
[0083]
In this method, when search processing is performed in a plurality of database systems, the search target documents are presented to the user in the order in which the search target documents are obtained, and the waiting time can be reduced.
[0084]
In addition, for each of one or more selected groups, the search unit 207 calculates a document score according to a search request that is accepted for a document included in the group, and calculates the calculated group score and The total score may be calculated with reference to the calculated document score, and the search results for the search request accepted for input may be output in the order of the total score.
[0085]
The document score can be obtained, for example, by calculating a cosine between a keyword vector and an input request vector included in the search target document. The total score is the linear sum of the group score and the document score Etc. Can be considered.
[0086]
In this method, the order in which the search results are presented to the user is appropriately determined using the total score, and the search results can be presented in the order according to the desire desired by the user.
[0087]
As described above, in this embodiment, the search target documents are classified into groups based on a predetermined criterion in advance, a group score corresponding to the criterion is calculated from the search request, and a group to be searched is selected by this, and the load is calculated. In addition to achieving dispersion, the user can obtain search results with a short waiting time.
[0088]
Also, the index creation process and the search response process can be executed independently. For example, an index creation process may be performed every predetermined period or when a search response process is not being executed (idle state waiting for a search request input from a user) for a predetermined time or longer. It is done.
[0089]
(Embodiment of grouping)
In the following, embodiments of grouping various search target documents will be described in detail. First, it is a classification method depending on the contents of a document.
[0090]
In the first method, N case documents to which one or more labels indicating classification categories are assigned are prepared in advance, and classification is performed according to the degree of similarity with the case documents. There are n types of labels.
[0091]
There are various methods for calculating the degree of similarity (score) between documents. As described above, using a keyword appearance frequency vector, the cosine (inner product) between them is used as the degree of similarity.
[0092]
The degree of similarity between a certain search target document and each case document is calculated, and this is used as a score to select the top M (M ≦ N) case documents in the score.
[0093]
Then, each score is added to the score for the label given to the case document. Since the same label may be assigned to different case documents, the score of the label may be a sum of scores of a plurality of case documents.
[0094]
Then, the search target document is classified into a classification category represented by a label having the highest score.
[0095]
The second method also uses a similar search. In this method, an appropriate search request is prepared for classification.
[0096]
Then, using the similar search, the score of each search target document for this search request is calculated, and the top one search target document is obtained.
[0097]
Next, the top first document is used as the next search request, and the score of each remaining search target document is calculated. This score of the search target document is called a primary score. And the top M of the primary score ₁ Pieces (M ₁ ≦ N) and select this as the first group.
[0098]
Similarly, next, a document whose primary score is second or lower and which does not belong to the first group is used as the next search request, and the score of each other search target document is calculated. Obtain secondary score, top M of secondary score ₂ Pieces (M ₂ ≦ N) and select this as the second group.
[0099]
Then, the common elements of the first and second groups are reclassified into the group with the higher score.
[0100]
Further, a search target document that does not belong to either the first group or the second group is selected, and the same processing is repeated to classify the documents.
[0101]
As described above, various document classification methods depending on the contents of the document can be applied to the present invention.
[0102]
Hereinafter, a classification method that does not depend on the contents of a document will be described in detail. This focuses on the apparent features of the document, not the content of the document. That is, classification is performed based on the following information.
-Document creation date, update date.
・ Document creator.
-Document management number.
Any explicit descriptor included in the document. For example, description of keywords in META tags in HTML documents.
-Where the document is located. URL or absolute path on the file system.
[0103]
A case where classification is performed based on a URL having a tree structure will be described. The tree structure can be divided into a plurality of groups by cutting it with a cross section having an arbitrary depth. Then, the number of search target documents belonging to the respective groups is checked, and if this is larger than a predetermined number, the document is re-divided in a deeper section. On the other hand, if it is smaller than the predetermined number, it is merged with any other group. This is repeated until the number of search target documents included in each group falls within a desired range.
[0104]
Thus, when classifying using such information, a large amount of documents can be easily classified.
[0105]
(Embodiment of calculation of group score)
In the following, embodiments of group score calculation will be described in detail.
[0106]
Before calculating the group score, first, as a preparation, a set of terms included in the input search request is configured. It is desirable to extend these terms using a thesaurus or normalize them as appropriate. Also, feature words may be extracted from the obtained search result document and added to the term set.
[0107]
When there are n groups, the method for calculating the group score score (t, c) of the group c for each term t includes, for example, the following.
[0108]
(1) DF x ICF ₁ What is called. The following calculation formula is used.
score (t, c) = DF × ICF
However,
DF = df _c / (df _c + K);
ICF = log ((n + 0.5) / CF) / log (n + 1.0);
K is a constant determined by empiricall y;
df _c Is the proportion of documents in group c in which the term t appears;
CF is the number of groups containing documents in which the term t appears
[0109]
(2) DF x ICF ₂ What is called. The following calculation formula is used.
score (t, c) = DF × ICF
However,
DF = dt + (1-dt) × df _c / max _c (df _c );
ICF = log (n / CF);
dt is a constant not less than 0 and not more than 1, typically 0.5;
df _c Is the proportion of documents in group c in which the term t appears;
max _c (df _c ) For each group df _c The maximum value of;
CF is the number of groups containing documents in which the term t appears
[0110]
(3) What is called DF × AVG-IDF. The following calculation formula is used.
score (t, c) = DF × AVGICF
However,
DF = dt + (1-dt) × df _c / max _c (df _c );
AVGICF = (1 / n) Σ _c log (n / df _c ) + 1;
dt is a constant not less than 0 and not more than 1, typically 0.5;
df _c Is the proportion of documents in group c in which the term t appears;
max _c (df _c ) For each group df _c The maximum value of
[0111]
With such a calculation formula, a score of group c for each term t is obtained. Therefore, the sum of the term scores included in the search request is set as the group score of the group c.
[0112]
In order to perform these calculations, the index unit 203 determines that “the total number of words included in the search target documents included in the group” “the number of occurrence documents of the term included in the group” “the number of groups including the document in which the term appears. It is necessary to memorize information such as “
[0113]
In addition, the following processing may be performed.
[0114]
DF × ICF ₁ Decrease the value of K in DF × ICF ₂ If the value of dt is decreased in DF × AVGICF, the score of a group having a large size (the number of documents included in the group is large) tends to increase. Therefore, these values are adjusted according to the user's instruction when calculating the group score depending on whether the reproduction rate is important or the initial accuracy is important.
[0115]
Further, the total score s obtained may be adjusted by the following calculation formula to be used as a group score.
s × (log (n) + k)
By adjusting the parameter k, the group score can be set to a desirable value.
[0116]
In addition, when the number of terms included in the search request is large, a method of setting score (t, c) = 0 without calculating for a term having a large CF can be considered. It is also effective to not register such terms as stop words in the index unit 203.
[0117]
Furthermore, a priority term is assigned to each group, and when the priority term is included in the search request, a mode in which the group is preferentially selected (with the highest score) is also conceivable.
[0118]
【The invention's effect】
As described above, according to the present invention, it is possible to provide a search device suitable for performing a high-speed search, a search method, and a program for causing a computer to realize them.
[Brief description of the drawings]
FIG. 1 is a schematic diagram showing a schematic configuration of a typical information processing apparatus that implements a sub apparatus constituting a search apparatus of the present invention.
FIG. 2 is a schematic diagram showing a schematic configuration of a search device according to the present embodiment.
FIG. 3 is a flowchart showing a flow of control of index creation processing according to the present embodiment.
FIG. 4 is a flowchart showing a control flow of search response processing according to the present embodiment.
[Explanation of symbols]
101 Information processing apparatus
102 CPU
103 ROM
104 HD
105 keyboard
106 mice
107 RAM
108 Display device
109 interface
110 FD drive
111 CD-ROM drive
201 Search device
202 Group division
203 Index
204 Input reception part
205 Group score calculator
206 Group selection part
207 Search part
208 annex

Claims (8)

A group division unit for dividing the search target document into a plurality of groups;
For each of the plurality of divided groups, an index that associates and stores information on the group, a keyword that appears in a search target document included in the group, and the number of search target documents in which the keyword appears And
An input receiving unit that receives an input of a search request;
For each of the plurality of divided groups, the number of keywords appearing in the search request for which the input has been accepted, or a vector whose elements are weighted to the number of the keywords, are stored in the index unit. A vector whose element is the number of documents to be searched for in which the keyword appears in the group, and a group score calculation unit which calculates a cosine of the calculated cosine value and a group score.
A group selection unit that selects one or more groups from the plurality of groups based on the calculated group score;
For each of the one or more selected groups, a search unit that performs a search in parallel according to a search request that has received the input;
Merging the results searched for each of the selected one or more groups, and outputting the result as a search result for the search request received the input;
With
The search unit includes, for each of the selected one or more groups, a vector whose elements are the number of keywords appearing in a search request for which the input is accepted for documents included in the group, and the document A vector whose element is the number of times the keyword appears in, and the cosine of the vector, the calculated cosine value as the document score,
The merging unit calculates a linear sum of the calculated group score and the calculated document score, sets this as a total score, and sets a search result for the search request accepted as the input as the total score. A search device that outputs in the order of scores. The search device according to claim 1,
Each search target document is associated with a URL (Universal Resource Locator),
The group dividing unit determines that the search target documents include a tree structure of URLs associated with the search target documents when the number of search target documents included in each group exceeds (a) the number of desired ranges. Subdivide the group by cutting at
(B) If the number is less than the number of desired ranges, merge the group and other groups so that the number of search target documents included in the merged group is less than or equal to the number of desired ranges. A search device characterized by. The search device according to claim 1 or 2,
The indexing unit further stores, for each of the plurality of divided groups, information on a group and the total number of words included in a search target document included in the group in association with each other. . The search device according to any one of claims 1 to 3,
For each of the plurality of divided groups, the index unit includes the keyword, the group, and the appearance of the keyword in the group for all keywords included in the search target document included in the group. A search device characterized by further storing information associated with the number of documents. The search device according to any one of claims 1 to 4,
The group dividing unit calculates, for each of the search target documents, a cosine of a vector composed of frequencies of a plurality of words appearing in the search target document and a predetermined plurality of vectors, and the search target document is determined as the predetermined plurality A search apparatus, wherein the search target document is classified into a group associated with a vector having the largest cosine among the vectors. The search device according to any one of claims 1 to 4,
  The group dividing unit calculates a document score for each of the search target documents according to a predetermined search request, and classifies the search target documents according to the calculation result.
  A search device characterized by that. A group division process for dividing the search target document into a plurality of groups;
  For each of the plurality of divided groups, an index that associates and stores information on the group, a keyword that appears in a search target document included in the group, and the number of search target documents in which the keyword appears Process,
  An input receiving process for receiving an input of a search request;
  For each of the plurality of divided groups, the number of keywords appearing in the search request for which the input has been accepted, or a vector whose elements are weighted numbers of the keywords, and stored in the indexing step A group score calculation step of calculating a cosine of a vector having the number of documents to be searched for in which the keyword appears in the group as an element, and using the calculated cosine value as a group score;
  A group selection step of selecting any one or more groups from the plurality of groups based on the calculated group score;
  For each of the one or more selected groups, a search step of performing a search in parallel by a search request that has received the input;
  Merging the search results for each of the selected one or more groups, and outputting this as a search result for the search request that received the input;
  With
  In the search step, for each of the selected one or more groups, a vector having as an element the number of keywords appearing in the search request that has received the input for the documents included in the group, and the document A vector whose element is the number of times the keyword appears in, and the cosine of the vector, the calculated cosine value as the document score,
  In the merging step, a linear sum of the calculated group score and the calculated document score is calculated, and this is used as a total score, and a search result for the search request that has received the input is calculated as the total score. Output in order of score
  A search method characterized by that. A program that causes a computer to function as the search device according to any one of claims 1 to 6.

Images