JP2016015179A - Search device and search method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a search device and a search method capable of returning an appropriate search result even when a negative value is used in a WAND algorithm.SOLUTION: A search device 10 for searching a desired document from a plurality of documents on the basis of queries includes: query acquisition means 131 for acquiring the queries and query scores; a storage part 12 for storing indexes where document scores indicating degrees of association between the documents and the queries are recorded for each query and document, and the minimum value of the document scores is recorded for each query; and search means 133 for searching the predetermined number of documents with respect to the queries on the basis of total scores calculated from the query scores and the document scores. The search means 133 determines the document as the target of the calculation of the total scores by using the query scores and the minimum value when the query scores are negative values.

Description

  The present invention relates to a search device and a search method using a WAND algorithm.

Conventionally, a WAND search algorithm is known as a search algorithm that returns a search result in response to a user query (see, for example, Non-Patent Document 1).
As described in Non-Patent Document 1, in the WAND search algorithm, each document has a score for a query, and the maximum score value is recorded in the index. Then, based on the query score set for the query and the maximum score recorded in the index, the search result of the search target number (Top-k) is obtained.
Here, the product of the score for the query set in the index and the query score set for the query is calculated for each query, and the calculated product is added to obtain the total score for the document. At this time, in the WAND algorithm, as described above, the maximum score for each query of the document is recorded in the index. Therefore, when the total score for one document is calculated, the total score for other documents is set with the total score as a threshold value. It can be determined whether or not the calculation of can be skipped.

  For example, the query score for the query A is “1”, the query score for the query B is “2”, and the documents 1 to 5 are scored for the query A and the score for the query B as shown in Table 1 below. Is set. For simplicity of explanation, k = 1 is set (one with the highest score is extracted).

In the above example, the maximum score value “4” for query A and the maximum score value “3” for query B are recorded as indexes.
In the WAND algorithm, the total score for queries A and B for document 1 is calculated as 1.multidot.2 + 2.multidot.1 = 4. Here, the threshold is set as 4. Since the maximum score value for the query A is “4” and the query score is “1”, the document including only the query A cannot exceed the threshold value “4”. Therefore, document 2 having only a score for query A is skipped.
On the other hand, since the maximum value for the query B is “3” and the query score is “2”, the document including only the query B may exceed the threshold “4”. However, in the document including only the query B, the threshold is not exceeded unless the score is 3 or more, so the document 3 is also skipped.
Since the document 4 has scores for the queries A and B and has a higher score than the document 1, it is determined that the document 4 is worth calculating the total score. As a result, a total score “7” is calculated for the document 4. Here, the threshold “4” is updated to “7”. Thereafter, in the same manner, skip determination is performed on the remaining documents, and if not skipped, a total score is calculated.

“Deepak Agarwal”, “Maxim Gurevich”, “Fast top-k retrieval for model based recommendation”, “WSDM '12 Proceedings of the fifth ACM international conference on Web search and data mining”, USA, “ACM New York”, 2012 February 8, 2012, pages 483-492

By the way, in recent years, in order to return a more optimal search result to the user, a negative value may be used as a query score or a document score for the query. For example, when query B is not desired to be included in the search, the score of query B is set to a negative value, and a document with a low score for query B is searched.
However, in the WAND algorithm as described above, an optimal search result may not be returned if a negative score is used.

  For example, the score for the query A is “1”, the score for the query B is “−1”, and the scores for the query A and the score for the query B are as shown in Table 2 below for the documents 1 and 2. Suppose that it is set. For simplicity of explanation, k = 1 is set (one with the highest score is extracted).

When the Top-1 search by the WAND algorithm is performed in the above example, the total score for the queries A and B is calculated as 1 · 3 + (− 1) · (−1) = 4 for the document 1. On the other hand, since the threshold value is “4”, in the conventional WAND algorithm, since there is no document that exceeds the maximum score “3” of query A and the maximum score “−1” of query B, the total score for document 2 is calculated. Will be skipped.
However, the score for the actual document 2 is 1 · 1 + (− 1) · (−6) = 7, which is higher than the document 1 and should be searched higher than the document 1 originally. It becomes.
Thus, when a negative value is included in the score, there is a problem that an appropriate search result cannot be returned by a search using the WAND algorithm.

  An object of the present invention is to provide a search device and a search method capable of returning an appropriate search result even when a negative value is used in the WAND algorithm.

  The search device of the present invention is a search device that searches a desired document from a plurality of documents based on a query, the query acquisition means for acquiring the query and a query score that is a weight of the query, and the document An index storage means for storing an index in which a document score indicating a degree of relevance to the query is recorded for each query and each document, and a minimum value of the document score is recorded for each query; and the query score And a search means for searching a predetermined number of documents for the query based on the total score calculated by the document score, and the search means, when the query score is a negative value, Using the minimum value, the total score Characterized in that to determine the subject of the document to be calculated.

  In the present invention, when the query score is a negative value, the search processing by the WAND algorithm is performed using the minimum value of the document score. Therefore, the document can be searched in an accurate rank with respect to the query, and an appropriate search result is obtained. Can return.

The block diagram which shows the outline of the search device of 1st embodiment. The flowchart which shows the search process (search method) of 1st embodiment. The flowchart which shows the skip determination process in FIG. The figure which shows the outline of the advertisement search system which concerns on 2nd embodiment.

[First embodiment]
Hereinafter, a search device according to an embodiment of the present invention will be described with reference to the drawings.
[Configuration of search device]
FIG. 1 is a block diagram showing the search device of this embodiment.
The search device 10 of this embodiment is a computer, and includes a communication unit 11, a storage unit 12, a control unit 13, and the like.
The communication unit 11 is connected to a network via, for example, a LAN and communicates with other devices on the network.

The storage unit 12 is a data recording device constituted by, for example, a memory, a hard disk, etc., and constitutes an index storage unit in the present invention.
The storage unit 12 stores various programs and various data for controlling the search device. In addition, the storage unit 12 functions as an index storage unit and stores an index of a document to be searched.
Note that the document to be searched may be stored in a predetermined server device on the network or may be stored in the storage unit 12 of the search device 10.

  This index is index information for quickly searching for an optimum document for a query requested by a user or the like (search conditions such as keywords), and data as shown in Table 3 is recorded, for example.

As shown in Table 3, the index includes a document ID, a document score, a Max score, and a Min score.
The document ID is identification data that identifies a document.
The document score is set for each predetermined query and indicates the degree of relevance to the query. For example, when a search keyword is set as a query, the number of keywords included in the document or a numerical value calculated based on the number of keywords is recorded as the document score.
Further, as shown in the document score for the query B, the document score may set a negative value. For example, with respect to the query “for men”, the document ranking of the document “for women” is set to a negative value, so that the search order can be lowered.
The Max score is the maximum document score for each query, and the Min score is the minimum document score for each query.
In addition, as the index, the location of the document identified by the document ID (for example, URL) may be recorded.

  The control unit 13 includes an arithmetic circuit such as a CPU and a storage circuit such as a RAM. The control unit 13 expands a program (software) stored in the storage unit 12 or the like in the RAM, and cooperates with the program expanded in the RAM. Various processes are executed. And the control part 13 functions as the query acquisition means 131, the index acquisition means 132, the search means 133, etc. as shown in FIG. 1 by performing the said various processes.

  The query acquisition unit 131 acquires, for example, a query transmitted from a terminal device (user terminal) on the network and a query score that is a weight value of the query. In this embodiment, an example in which a query and a query score are acquired from a terminal device on the network is shown, but the present invention is not limited to this. For example, the search device 10 may include an input device such as a keyboard and a mouse, and a user of the search device 10 may input a query and a query score via the input device. In addition, the query and the query score are stored in the storage unit 12, and the query acquisition unit 131 acquires the query and the query score stored in the storage unit 12 according to a query designation request from a terminal device or an input device, for example. It is good also as a structure.

  The index acquisition unit 132 acquires an index stored in the storage unit 12. In this embodiment, an example in which an index is stored in the storage unit 12 is shown. However, for example, the index may be obtained from another server device on the network.

  Based on the query, the query score, and the index, the search unit 133 searches for an optimal document for the query using a WAND algorithm. Specifically, the search unit 133 functions as a score calculation unit 134, a threshold setting unit 135, a skip determination unit 136, and an extraction unit 137.

The score calculation unit 134 calculates a score (total score) for a predetermined query of a predetermined document based on the query score and the document score.
The threshold value setting unit 135 sets a threshold value for performing a search process using the WAND algorithm.
The skip determination unit 136 determines whether or not the total score of the document is calculated by the score calculation unit 134 based on the Max score, Min score, query score, and threshold value of the index.
The extracting unit 137 extracts the document ID to be searched for Top-k based on the calculated total score. That is, the top k document IDs with the highest total scores are extracted as search results.
A detailed description of each functional configuration will be described later.

[retrieval method]
Next, a Top-k document search method (search process) in the search apparatus 10 as described above will be described with reference to the drawings. FIG. 2 is a flowchart of search processing in the present embodiment.
As shown in FIG. 2, in the search processing based on the WAND algorithm in the present embodiment, first, the query acquisition unit 131 uses a query Q = {q ₁ , q ₂ , q ₃ ... Q _N } and these queries Q. Query score S _Q = {S _Q1 , S _Q2 , S _Q3 ... S _QN } is acquired (step S1).
In step S1, the query acquisition unit 131 may acquire a query Q and a query score S _Qn input from a terminal device such as a user terminal via a network, for example, as described above. According to the request, the query Q and the query score S _Qn stored in the storage unit 12 may be acquired.

  Thereafter, the index acquisition unit 132 reads the index stored in the storage unit 12 (step S2). In the present embodiment, the index stored in the storage unit 12 is read. However, as described above, the index may be acquired via a network.

Thereafter, various parameters in the search process are initialized (step S3).
Here, in the following description, it is assumed that the parameter relating to the document ID is i, the parameter relating to the number of documents (k) to be obtained from the search is j, and the threshold is X. In step S3, each parameter is initialized with i = 1, j = 0, and X = 0.

Thereafter, the score calculation means 134 of the search means 133 calculates the total score S _T (i, Q) for the document ID (denoted as DocID in FIGS. 2 and 3) i by the following equation (1) (step S4). ). In Expression (1), S _D (i, q _n ) is a document score for a query q _{n of} a document whose document ID is i, and S _Qn is a query score of the query q _n .

Next, the search unit 133 adds “1” to the parameter j (step S5), and determines whether or not the parameter j has reached the number of documents (k) to be obtained by the search process (step S6).
If it is determined as “No” in step S6, the search unit 133 adds “1” to the parameter i (step S7), and returns to the process of step S4.

If it is determined as “Yes” in step S6, the threshold setting unit 135 sets the threshold X in the WAND algorithm (step S8). Specifically, the threshold setting unit 135 sets the minimum value of the total scores S _T (i, Q) for k documents calculated by repeating Steps S4 to S7 as the threshold X.
After step S8, “1” is added to the parameter i (step S9), and a skip determination process (step S10) is performed.

[Skip judgment processing]
FIG. 3 is a flowchart of the skip determination process.
As shown in FIG. 3, in the skip determination process, the skip determination unit 136 first initializes a parameter n related to a query (n = 1) (step S101).
Next, the skip determination means 136 determines whether or not the query score S _Qn of the query q _n is a negative value (less than 0) (step S102).
If it is determined as “Yes” in step S102, the skip determination unit 136 uses the document score Min score S _Di (Min) for the query q _n of the document ID: i and the query score S _Qn of the query q _n. Then, a partial score S _n (i, q _n ) for the query q _n of the document is calculated by the following equation (2) (step S103).

On the other hand, in step S102, if it is determined as "No", the query score _{S Qn} query _{q n} determines whether it is positive value (greater than 0) (step S104).
If it is determined as “Yes” in step S104, the skip determination unit 136 uses the document score Max score S _Di (Max) for the query q _n of the document ID: i and the query score S _Qn of the query q _n. Then, a partial score S _n (i, q _n ) for the query q _n of the document is calculated by the following equation (3) (step S105).

On the other hand, if “No” is determined in step S104, the skip determining unit 136 determines that there is no document score (0) for the query q _n of the document ID: i. In this case, the skip determination unit 136 sets the partial score S _n (i, q _n ) for the query q _n of the document ID: i to 0 (step S106).

Thereafter, the skip determination unit 136 determines whether or not the parameter n is equal to or greater than the maximum number N of queries (step S107).
If it is determined as “No” in step S107, the skip determination means 136 adds “1” to the parameter n (step S108), and returns to step S102. That is, a partial score S _n (i, q _n ) of the document ID: i for each query Q acquired in step S1 is acquired.
If it is determined as “Yes” in step S107, the skip determination unit 136 adds the partial scores S _n (i, q _n ) calculated for each query q _n as shown in the following equation (4). Then, a provisional score _SF (i, Q) for the document ID: i is calculated (step S109).

Then, the skip determination unit 136 determines whether or not the calculated provisional score S _F (i, Q) is larger than the threshold value X (step S110). When it is determined “No” in step S110, the skip determination unit 136 sets document ID: i as a skip target (step S111). If it is determined as “Yes” in step S110, the skip determination unit 136 does not set the document ID: i as a skip target (is set as an evaluation target) (step S112).

  The skip determination process will be described using specific examples shown in Table 4 below.

In this example, k = 1. Further, it is assumed that the query score for the query A is “1” and the query score for the query B is “−1”.
In the above example, the total score S _T (1, Q) for the document 1 is calculated as 1 · 3 + (− 1) · (−1) = 4, and the threshold is set to “4”.
Since the query score of query A is a positive value for document 2, skip determination means 136 calculates a partial score for query A as 1 · 3 = 3 using Max score “3”. Since the query score of B is a negative value, (−1) · (−6) = 6 is calculated using the Min score “−6”. Therefore, the provisional score for the document 2 is calculated as S _F (2, Q) = 9. Thereby, the document 2 has a provisional score exceeding the threshold “4”, and is determined as an evaluation target (not skipped).

Returning to FIG. 2, after the skip determination process in step S10, the search unit 133 determines whether or not the document ID: i is a skip target by the skip determination unit 136 (step S11).
When it is determined as “No” in Step S11 (when it is determined to be evaluated), the score calculation unit 134 performs the same processing as in Step S4, and the total score S _T (i, Q) for the document with the document ID: i. Is calculated (step S12).
Then, the threshold setting unit 135 determines whether or not the total score S _T (i, Q) calculated in step S12 is larger than the threshold X (step S13). Set (step S14). That is, the threshold setting means 135 extracts the upper k pieces from the calculated total score S _T (i, Q), and sets the lowest value as the threshold X.

On the other hand, when it is determined as “Yes” in Step S11 (when it is determined as a skip target), the processes of Steps S11 to S14 are skipped.
That is, if “Yes” is determined in step S11, “No” is determined in step S13 (when the total score S _T (i, Q) does not exceed the threshold value X), and the search is performed after step S14. The means 133 adds “1” to the parameter i (step S15) and determines whether the parameter i exceeds the maximum value I (step S16). That is, it is determined whether or not the processing in steps S1 to S14 has been performed on all documents. If it is determined “No” in step S16, the process returns to step S10.

If it is determined as “Yes” in step S16, the extraction unit 137 extracts the top k out of the calculated total scores S _T (i, Q), and extracts the extracted total scores S _T (i, Q). ) As a search result (step S17).

[Operational effects of this embodiment]
In the present embodiment, the Max score and Min score of the document score for each query are recorded as indexes.
Then, as shown in steps S102 to S103, the search unit 133 uses the Min score when the query score is negative, and uses the Max score when the query score is positive. The search process based on
In general, when the WAND algorithm is used by using only the Max score, there is a possibility that the calculation of the total score should be skipped even for a document that should not be skipped. On the other hand, in this embodiment, it is possible to accurately determine whether or not the calculation of the total score may be skipped by performing a search process based on the WAND algorithm using the above-described index. Therefore, the search result of the document can be returned with an accurate ranking with respect to the query.

[Second Embodiment]
Next, 2nd embodiment in this invention is described based on drawing.
In the present embodiment, an advertisement search process will be described as an example of a search process using the search device shown in the first embodiment.
FIG. 4 is a diagram showing an outline of an advertisement search system that performs advertisement search. In addition, about the structure similar to 1st embodiment, the same code | symbol is attached | subjected and the description is abbreviate | omitted or simplified.
As illustrated in FIG. 4, the advertisement search system 100 includes the search device 10 described above, a terminal device operated by a user (user terminal 20), a terminal device operated by an advertiser (advertiser terminal 30), These devices 10, 20, and 30 are communicably connected via a network.

[User terminal]
Although not shown, the user terminal 20 is a computer, and includes an input device configured with a keyboard and the like, a storage device configured with a memory and the like, an arithmetic device configured with a CPU and the like, and a display.
Then, the user terminal 20 acquires user data when the user operates the input device, and transmits the acquired user data to the search device 10 as a query.
Here, the user data may be, for example, user personal data such as gender, age, residence area, etc., and it is desired to receive conditions for receiving desired advertisement distribution such as user interest data, and advertisement distribution. There may be no conditions.

In addition, the user terminal 20 obtains a priority for user data distributed as a query as a query score when the user operates the input device, and transmits the query score to the search device 10.
For example, it is possible to set a positive priority for a condition for receiving advertisement distribution and to set a negative priority for data for which advertisement distribution is not desired.
An example of the user data (query) and query score as described above is shown in Table 5.

[Advertiser terminal]
The advertiser terminal 30 is a computer and includes an input device configured with a keyboard and the like, a storage device configured with a memory and the like, an arithmetic device configured with a CPU and the like, and a display (not shown).
The storage device of the advertiser terminal 30 stores an advertisement (advertisement data) that is a document in the present invention to be distributed to the user terminal 20. Moreover, the advertiser terminal 30 acquires advertisement delivery data, for example, when an advertiser operates an input device. This advertisement distribution data is data in which a distribution condition corresponding to user data (query) and a score (that is, document score) indicating the degree of association of the advertisement with the distribution condition are associated with each other. Then, the advertiser terminal 30 transmits the acquired advertisement distribution data as a document score to the search device.
Note that the advertiser terminal 30 may transmit the advertisement stored in the storage device to the search device 10.
Table 6 shows an example of advertisement distribution data transmitted from the advertiser terminal 30 to the search device 10.

[Search device]
The search device 10 has substantially the same configuration as that of the first embodiment described above. The control unit 13 of the search device 10 functions as a query acquisition unit 131, an index acquisition unit 132, and a search unit 133 by reading and executing a program stored in the storage unit 12.
The query acquisition unit 131 acquires user data and query scores as shown in Table 5 described above from the user terminal 20 and stores them in the storage unit 12.
The index acquisition unit 132 acquires advertisement distribution data as shown in Table 6 described above from the advertiser terminal 30. Then, the index acquisition unit 132 integrates the advertisement distribution data transmitted from the plurality of advertiser terminals 30, newly reassigns the advertisement ID as the document ID, and stores the advertisement ID in the storage unit 12 as an index. At this time, the index acquisition means 132 stores the maximum document score for each query (user data) in the index as the Max score and the minimum value as the Min score. When new advertisement distribution data is acquired, the acquired advertisement distribution data is added to the index, and the Max score and Min score are recalculated.

Thereby, the search device 10 can execute the search processing based on the WAND algorithm similar to the first embodiment described above using the query, the index associated with the user data, the user data, and the index. An optimal document ID can be extracted for user data (query).
The search device 10 transmits the URL of the distribution source (advertiser terminal 30) of the advertisement data corresponding to the extracted document ID to the user terminal 20. When advertisement data is stored in the storage unit 12, the search device 10 reads out advertisement data corresponding to the extracted document ID from the storage unit 12 and transmits it to the user terminal 20. Thereby, optimal advertisement data can be distributed to each user terminal 20 for the user.

[Modification]
In 2nd embodiment, although the search apparatus 10 acquired the advertisement delivery data from the advertiser terminal 30 via a network, the example which produces an index based on the acquired advertisement delivery data was shown, for example, the search apparatus 10 An index may be created by operating.

  As an application example of the search device 10 of the first embodiment, an advertisement distribution system has been illustrated, but is not limited thereto. The present invention can be applied to an apparatus or a system that performs a search process using a query. For example, the present invention is applied to various search uses such as a search for contents and files based on a query such as a search keyword and a facility search for a store. be able to.

  In addition, the specific structure and procedure for carrying out the present invention can be appropriately changed to other structures and the like within a range in which the object of the present invention can be achieved.

  DESCRIPTION OF SYMBOLS 10 ... Search apparatus, 12 ... Memory | storage part (index storage means), 13 ... Control part, 20 ... User terminal, 30 ... Advertiser terminal, 100 ... Advertisement search system, 131 ... Query acquisition means, 132 ... Index acquisition means, 133 ... Search means, 134 ... Score calculation means, 135 ... Threshold setting means, 136 ... Skip determination means, 137 ... Extraction means.

Claims (5)

A search device for searching a desired document from a plurality of documents based on a query,
Query acquisition means for acquiring a query score that is a weight of the query and the query;
An index storage unit that stores an index in which a document score indicating a degree of relevance of the document to the query is recorded for each query and each document, and in which a minimum value of the document score is recorded for each query;
Search means for searching a predetermined number of documents for the query based on a total score calculated from the query score and the document score;
The search device, comprising: when the query score is a negative value, using the query score and the minimum value to determine a target document for calculating the total score. The search device according to claim 1,
The maximum value of the document score for each query is recorded in the index,
The search device, when the query score is a positive value, uses the query score and the maximum value to determine a target document for calculating the total score. The search device according to claim 2, wherein
For each document, the search means sets “0” when there is no document score, a product of the query score and the maximum value when the query score is positive, and a negative value for the query score. In some cases, the product of the query score and the minimum value is calculated as a partial score for the query, and when the sum of the partial scores calculated for all queries is equal to or less than a predetermined threshold, the document The search device characterized by skipping the calculation of the total score. The search device according to any one of claims 1 to 3,
The query is user data about the user;
The document is advertising data;
The index storage means stores, as the index, advertisement distribution data in which the relevance of the user data to the advertisement data is the document score. A search method for searching a desired document from a plurality of documents based on a query by a computer,
The computer stores an index in which a document score indicating a degree of relevance of the document to the query is recorded for each query and each document, and an index in which a minimum value of the document score is recorded for each query is stored. Having means,
In the search method, the computer includes:
Obtaining a query score that is a weight of the query and the query;
Based on the total score calculated by the query score and the document score, a predetermined number of documents for the query are searched,
In the calculation of the total score, when the query score is a negative value, a document for which the total score is calculated is determined using the query score and the minimum value.

Images