JP6789253B2 - Search device, search method, and program - Google Patents

Description

The present invention relates to a search device, a search method, and a program, and more particularly to a search device, a search method, and a program for improving the accuracy of reranking.

In search, the user of the search inputs a query, and the search algorithm calculates the measure of the relationship between the query and the data of the search target stored in the database, and the data of the search target with the highest relationship measure. It is to return to the user the data of the search target whose relationship measurement is higher than a predetermined threshold, the ranking list in which the search target data is sorted in descending order of the relationship measurement.

The search target includes, for example, documents, images, sounds, videos, various other media, and data recorded as a combination thereof. Hereinafter, the data to be searched is simply referred to as a search target. In addition, the measure of the relationship is described as the similarity, and the similarity between the query and the search target is also described as the rank score of the search target in the query.

Search targets that are related to the query for the user are described as related search targets, and search targets that are not related are described as irrelevant search targets.

The rank of the search target in the ranking list is called the rank of the search target. The rank value is a positive integer, which is a different concept from the rank score. High rank means that the rank value is small. For example, the search target having the highest rank has a rank value of 1.

However, when the amount of information obtained from the query is small, the similarity between the query and the related search target is low, and the related search target is likely to be given a low rank.

In order to deal with such a problem, various search target reranking methods for assigning a new rank to the search target given the rank by the first search have been proposed.

Below, among the search targets given a rank by the first search, those with a higher rank are described as high-ranking results. Among the search targets included in the top results, those that are related are described as true positive top results, and those that are not are described as false positive top results.

For example, the search target reranking device described in Non-Patent Document 1 uses a first query as an input to perform a first search for all search targets, and acquires a plurality of high-ranking results from the search targets.

Next, for each of the search targets, the search target is set as the second query, the second search is performed on the set of data that combines the first query and all search targets, and the first query and all search targets are selected. Give a rank.

Then, for each of the search targets, the rank score of the search target is calculated based on the ranks given to the first query and all the top results. After the calculation of the rank score for all the search targets is completed, all the search targets are sorted in descending order of rank score.

Further, for example, the search target reranking device described in Non-Patent Document 2 uses the first query as an input to perform the first search for all the search targets and assigns the first rank to all the search targets. , Get multiple top results from the search target.

Next, for each of the top results, the top result is set as the second query, the second search is performed on the set of data that combines the first query and all search targets, and the first query and all search targets are selected. Gives the second rank.

Then, for each of the search targets, the rank of the search target is based on the first rank matrix of all the high-ranking results, the second rank matrix of the first query, and the second rank matrix of the search target. Calculate the score.

After the calculation of the rank score for all the search targets is completed, all the search targets are sorted in descending order of rank score.

Further, for example, the search device described in Non-Patent Document 3 has similarities based on the distance between two data constituting the pair for each pair of data in a set of data in which a query and all search targets are combined. Calculate the degree and determine the initial value of the rank score.

Next, after the calculation of the similarity for all the pairs of data is completed, the rank score of the data is updated for each of the data based on the similarity between the data and all the data and the rank score of all the data.

Then, the rank score is updated for all of the data, the update process is repeated a plurality of times, and all the search targets are sorted in descending order of the rank score obtained in the last update.

Danfeng Qin, Stephan Gammeter, Lukas Bossard, Till Quack, and Luc J. Van Gool. Hello neighbor: Accurate object retrieval with k-reciprocal nearest neighbors. In CVPR, 2011, pp. 777-784. Xiaohui Shen, Zhe Lin, Jonathan Brandt, and Ying Wu. Spatially-constrained similarity measure for large-scale object retrieval. IEEE Trans. Pattern Anal. Mach. Intell., Vol. 36, No. 6, 2014, pp. 1229- 1241. Dengyong Zhou, Jason Weston, Arthur Gretton, Olivier Bousquet, and Bernhard Scholkopf. Ranking on data manifolds. In NIPS, 2003, pp. 169-176.

The accuracy of the search target reranking method strongly depends on the accuracy of the contribution of the top results to the reranking.

When the contribution of false positive top results to the reranking is calculated high, the unrelated search target having a high degree of similarity to the top results is calculated with a high rank score, which adversely affects the accuracy of the reranking.

For example, the method described in Non-Patent Document 1 uniformly treats the degree of contribution to the reranking of all top results. Further, in the method described in Non-Patent Document 2, a weighted value inversely proportional to the first rank of each higher-ranking result is calculated as a degree of contribution to the re-ranking of the higher-ranking result.

However, the accuracy of the first rank strongly depends on the accuracy of the first search, and when the accuracy of the first search is low, the first rank is high and a high weighted value is given to the top result of false positive. Probability is high.

Such high-ranked false-positive results of the first rank still have a high degree of contribution to re-ranking, and often adversely affect the accuracy of re-ranking.

Therefore, even if the degree of contribution to the re-ranking of the false positive top results and the degree of similarity with the search target unrelated to the true positive top results are low, the degree of similarity is judged to be high. , There was a problem that the accuracy of reranking deteriorated.

The present invention has been made in view of the above points, and an object of the present invention is to provide a search device, a search method, and a program capable of improving the accuracy of reranking and improving the accuracy of search.

The search device according to the present invention uses the data as a second query for each of the data included in the data set including the input first query data and the plurality of data to be searched as the elements. For each of the search unit that assigns a rank that is the rank related to the second query of each of the data included in the data set when the search is performed on the set, and each of the data pairs included in the data set. , When the similarity between the one data and the other data is the rank given to the other data when one data is used as the second query, and when the other data is used as the second query. A similarity calculation unit that calculates so that the higher the rank of the rank assigned to one of the data in the above, the larger the value, and the similarity calculation unit for each of the plurality of data to be searched. A rank score calculation unit for calculating a rank score indicating the degree to which the data is related to the first query is provided based on the similarity obtained by the above.

Further, in the search method according to the present invention, the search unit obtains the data for each of the data included in the data set including the input first query data and the plurality of data to be searched as elements. When a search is performed on the data set as the second query, a search step for assigning a rank, which is a rank related to the second query, for each of the data included in the data set, and a similarity calculation unit are used. For each of the pairs of data included in the data set, the similarity between the one data and the other data is the rank given to the other data when one of the data is used as the second query. , The similarity calculation step that calculates so that the higher the rank, the larger the rank given to one of the data when the other data is used as the second query, and the rank score calculation unit , A rank score calculation step for calculating a rank score indicating the degree to which the data is related to the first query based on the similarity obtained by the similarity calculation unit for each of the plurality of data to be searched. And, including.

According to the search device and the search method according to the present invention, the search unit determines each of the data included in the data set containing the input first query data and the plurality of data to be searched as elements. When a search is performed on the data set using the data as the second query, each of the data contained in the data set is given a rank which is the rank related to the second query, and the similarity calculation unit gives the data set a rank. For each of the included data pairs, the similarity between one data and the other data is the rank given to the other data when one data is used as the second query, and the other data. When the data is used as the second query, the higher the rank, the larger the value given to the data.

Then, the rank score calculation unit calculates a rank score indicating the degree to which the data is related to the first query based on the similarity obtained by the similarity calculation unit for each of the plurality of data to be searched. ..

As described above, for each of the data pairs included in the data set whose elements are the input first query data and the plurality of data to be searched, one of the relevant data and the other of the relevant data The degree of similarity is determined by the rank given to the other data when one of the data is used as the second query and the rank given to one of the data when the other data is used as the second query. By calculating so that the higher the rank of both, the higher the value, the similarity between the related search target and the unrelated search target is suppressed to a low level, and the rank score of the unrelated search target is calculated more accurately. , The accuracy of re-ranking can be improved and the accuracy of search can be improved.

Further, the similarity calculation unit of the search device according to the present invention determines the similarity between one of the data and the other of each of the pairs of data included in the data set, and the second of the data. The rank given to the other data in the case of a query and the rank given to one of the data in the case of using the other data as a second query are both higher ranks than a predetermined rank. In both cases, the higher the rank, the larger the value, and when at least one of the ranks is lower than the predetermined rank, the smaller the value can be calculated.

Further, in the similarity calculation step of the search method according to the present invention, for each of the pairs of data included in the data set, the similarity between one of the data and the other of the data is obtained, and one of the data is used. The rank given to the other data in the case of two queries and the rank given to one of the data in the case of using the other data as the second query are both higher than the predetermined rank. In some cases, the higher the rank, the larger the value, and when at least one of the ranks is lower than the predetermined rank, the smaller the value can be calculated.

In addition, the search device according to the present invention obtains the data for each of the data included in the first data set whose elements are the input first query data and the plurality of data to be searched. When a search is performed on the first data set as a query, a search unit that assigns a rank that is a rank related to the second query of each of the data included in the first data set, and the first query. For each of the pair of data included in the second data set whose elements are the data having the highest rank N and the data to be searched, the degree of similarity between the one data and the other data is determined. , The rank given to the other data when one of the data is used as the second query and the rank given to one of the data when the other data is used as the second query are both high. For each data of the first similarity calculation unit and the second data set, which is calculated so as to take a larger value as the rank is, the data is based on the similarity obtained by the first similarity calculation unit. The first rank score calculation unit that calculates the first rank score indicating the degree of relevance to the first query, each of the data included in the plurality of data to be searched, and the data included in the second data set. For each of the pairs in combination with each of the above, a second similarity, which is the similarity between the one data and the other data, is given to the other data when one of the data is used as the second query. A second similarity calculation unit that calculates so that the higher the rank, the larger the value obtained and the rank given to one of the data when the other data is used as the second query. For each of the plurality of data to be searched, a second rank score indicating the degree to which the data is related to the first query is obtained based on the first rank score of the data and the second similarity. It is configured to include a second rank score calculation unit for calculating.

Further, in the search method according to the present invention, the search unit describes each of the data included in the first data set including the input first query data and the plurality of data to be searched as elements. When a search is performed on the first data set using the data as the second query, a search step for assigning a rank, which is a rank related to the second query, for each of the data included in the first data set, and a search step. One of the first similarity calculation units is for each of the data that is the first query and the data pair that is included in the second data set that includes the data to be searched for the top N ranks. The degree of similarity between the data and the other data is the rank given to the other data when one of the data is used as the second query, and one of the ranks given to the other data when the other data is used as the second query. The first similarity calculation step, which calculates so that the higher the rank given to the data, the larger the value, and the first rank score calculation unit determines each data in the second data set. A first rank score calculation step for calculating a first rank score indicating the degree to which the data is related to the first query based on the similarity obtained by the first similarity calculation unit, and a second similarity calculation. For each of the pairs of data included in the plurality of data to be searched and each of the data included in the second data set, one of the data and the other data The second similarity, which is the similarity, is set to the rank given to the other data when one of the data is used as the second query and to one of the data when the other data is used as the second query. The second similarity calculation step, which calculates so that the higher the assigned rank, the larger the value, and the second rank score calculation unit describes each of the plurality of data to be searched. Includes a second rank score calculation step of calculating a second rank score indicating the degree to which the data is relevant to the first query, based on the first rank score of the data and the second similarity.

According to the search device and the search method according to the present invention, each of the data included in the first data set whose elements are the input first query data and the plurality of data to be searched by the search unit. When a search is performed on the first data set using the data as the second query, each of the data included in the first data set is given a rank, which is the rank related to the second query, and the first For each of the pair of data included in the second data set whose elements are the data that is the first query and the data to be searched in the top N ranks, the similarity calculation unit has one of the relevant data and the other of the data. The degree of similarity with the data is given to the rank given to the other data when one of the data is used as the second query and to one of the data when the other data is used as the second query. The higher the rank, the larger the value is calculated so that the first rank score calculation unit obtains the similarity obtained by the first similarity calculation unit for each data in the second data set. Based on this, a first rank score indicating the degree to which the data is relevant to the first query is calculated.

Then, the second similarity calculation unit uses one of the data and the other for each of the pairs of the data included in the plurality of data to be searched and the data included in the second data set. The second similarity, which is the similarity with the relevant data, is the rank given to the other relevant data when one of the relevant data is used as the second query, and the case where the other relevant data is used as the second query. One of the ranks assigned to the data is calculated so that the higher the rank, the larger the value, and the second rank score calculation unit calculates the first of the data for each of the plurality of data to be searched. Based on the 1st rank score and the 2nd similarity, the 2nd rank score indicating the degree to which the data is related to the 1st query is calculated.

In this way, for each of the data included in the first data set whose elements are the input first query data and the plurality of data to be searched, the data is used as the second query in the first data set. Each of the data included in the first data set is given a rank which is the rank related to the second query, and the first similarity calculation unit is the data which is the first query. For each of the data pairs included in the second data set whose elements are the data to be searched with the top N ranks, the similarity between the data on one side and the data on the other side is determined. The higher the rank, the larger the value given to the other data in the case of the second query and the rank given to one of the data in the case of using the other data as the second query. For each data in the second data set, the first rank score indicating the degree to which the data is related to the first query is calculated based on the similarity, and the data to be searched is divided into a plurality of data to be searched. For each pair of combinations of each of the included data and each of the data contained in the second data set, the second similarity, which is the similarity between the one data and the other data, is the one. The higher the rank given to the other data when the data is used as the second query, and the rank given to one of the data when the other data is used as the second query. Calculated to take a large value, and for each of the plurality of data to be searched, the degree to which the data is related to the first query is shown based on the first rank score and the second similarity of the data. By calculating the second rank score, the degree of contribution to the reranking of the top false positive results and the similarity between the top results of true positives and unrelated search targets are suppressed to a low level, and the accuracy of reranking is improved. , The accuracy of the search can be improved.

Further, the first similarity calculation unit of the search device according to the present invention determines the similarity between one of the data and the other of each of the data pairs included in the second data set. The rank given to the other data when the data is used as the second query and the rank given to one of the data when the other data is used as the second query are both higher than the predetermined ranks. When the rank is high, the higher the rank, the higher the value, and when both ranks higher than the predetermined rank, the higher the rank, the higher the value. It is calculated so as to take a small value when the rank is lower than the predetermined rank, and the second similarity calculation unit includes each of the data included in the plurality of data to be searched and the second data set. For each of the pairs that combine each of the data, the second similarity, which is the similarity between the one data and the other data, is used as the second query, and the other data is used as the second query. When the rank given to the data and the rank given to one of the data when the other data is used as the second query are both higher than the predetermined rank, the higher the rank, the larger the rank. It can be calculated to take a value, and if at least one is a rank lower than the predetermined rank, it can be calculated to take a smaller value.

The program according to the present invention is a program for functioning as each part of the above-mentioned search device.

According to the search device, search method, and program of the present invention, the accuracy of reranking can be improved to improve the accuracy of search.

It is the schematic which shows the structure of the search apparatus which concerns on 1st Embodiment of this invention. It is a flowchart which shows the search processing routine of the search apparatus which concerns on 1st Embodiment of this invention. It is the schematic which shows the structure of the search apparatus which concerns on 2nd Embodiment of this invention. It is a flowchart which shows the search processing routine of the search apparatus which concerns on 2nd Embodiment of this invention.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.

<Overview of the search device according to the first embodiment of the present invention>
First, the outline of the first embodiment of the present invention will be described.

In the method described in Non-Patent Document 3, the accuracy of the search strongly depends on the accuracy of the similarity between all the data required for updating the rank score. The method uses similarity based on the distance between data, but if the distance calculation is not accurate, it is likely that the similarity between the related search target and the unrelated search target is high. ..

For this reason, there is a problem that the rank score of an irrelevant search target is calculated to be excessively high, which often adversely affects the accuracy of the search.

In the present embodiment, for each of the data included in the data set containing the input first query data and the plurality of data to be searched as elements, the data is used as the second query for the data set. Search and give a rank to all data.

Then, for each of the data pairs, the similarity between the one data and the other data is the rank given to the other data when one data is used as the second query, and the other data. When the data is used as the second query, the higher the rank, the larger the value given to the data.

Then, after the calculation of the similarity for all the data pairs is completed, the rank score of all the data is calculated based on the calculated similarity between all the data.

With such a configuration, it is possible to suppress the similarity between the related search target and the unrelated search target to a low level, calculate the rank score of the unrelated search target more accurately, and improve the accuracy of the search.

Further, for each of the data pairs, the similarity is defined as the rank given to the other data when one of the data is used as the second query and the other when the other data is used as the second query. When the ranks given to the relevant data are both higher than the predetermined rank, the higher the rank, the larger the value is calculated, and at least one of the ranks is lower than the predetermined rank. It may be calculated to take a small value in some cases.

With such a configuration, the degree of similarity between the related search target and the unrelated search target can be suppressed to be lower, and the accuracy of the search can be further improved.

<Structure of the search device according to the first embodiment of the present invention>
The configuration of the search device according to the first embodiment of the present invention will be described with reference to FIG. FIG. 1 is a block diagram showing a configuration of a search device according to the present embodiment.

The search device 10 is composed of a computer including a CPU, a RAM, and a ROM that stores a program for executing a search processing routine described later, and is functionally configured as shown below.

As shown in FIG. 1, the search device 10 according to the present embodiment includes an input unit 100, a database 110, a search unit 120, a similarity calculation unit 130, a rank score calculation unit 140, and an output unit 150. Be prepared.

The input unit 100 receives the input of the first query q, which is a search query. Then, the input unit 100 passes the received first query to the search unit 120.

The database 110 stores a plurality of data to be searched.

Specifically, the database 110 is the data to be searched.

Is memorized n. The database 110 is the data to be searched in response to the request of the search unit 120.

Is passed to the search unit 120.

The search unit 120 searches the data set using the data as the second query for each of the data included in the data set including the input first query data and the plurality of data to be searched. When this is done, each of the data contained in the data set is given a rank, which is a rank related to the second query.

Specifically, when the search unit 120 first receives the first query q from the input unit 100, the search unit 120 receives a plurality of data to be searched from the database 110.

Is obtained, and these are set as one set.

To create. here,

And.

Next, the search unit 120 sets data.

For each of the data contained in the data, when the data is searched as the second query for each of the data included in the data set, the pair of the second query and the target data is referred to. Give a rank, which is the rank related to the second query.

More specifically, first, x ₀ is used as the second query, and the data set

Search for all the data included in, and give a rank to all the data. Here, the rank given to the target data x _j when the data x _i is used as the second query is expressed as r (x _i | x _j ).

For example, if the rank of the searched second query x ₀ and the target data x 2 is 3, then r (x ₀ | x ₂ ) = 3.

The search unit 120 refers to each of the pair {x ₀ , x ₀ }, {x ₀ , x ₁ }, ..., {x ₀ , x _n } of the data for which the second query is targeted as x ₀ . Ranks are given such as r (x ₀ | x ₀ ), r (x ₀ | x ₁ ), ..., r (x ₀ , x _n ), and the like.

Then, the search unit 120 similarly sets the data set when the second query is x ₁ , when the second query is x _2, and when all the data in the data set is the second query.

A rank is given to all the data contained in.

Here, as the search method, for example, when the search target is a document, the bag-of-words method, etc., and when the search target is an image, the bag-of-visual-words method, etc., depending on the nature of the search target. Various search methods can be used.

Next, when the search unit 120 assigns ranks to all the data pairs, it generates a rank matrix that stores the values of all the obtained ranks, and the rank matrix and the generated data set are generated.

Is passed to the similarity calculation unit 130.

For each pair of data included in the data set, the similarity calculation unit 130 sets the similarity between one of the data and the other of the data as the second query of the other data. It is calculated that the higher the rank, the larger the value given to the data and the rank given to one of the data when the other data is used as the second query.

Specifically, the similarity calculation unit 130 sets data based on the rank matrix.

For each of the data pairs included in, the similarity s (x _i , x _j ) is obtained for the data pair {x _i , x _j }.

Here, the similarity calculation unit 130 assigns the similarity s (x _i , x _j ) to the rank r (x _i |) given to the other data x _j when one data x _i is used as the second query. to take a large value as _{x i)} and is at both high rank _| and x _j), the rank given to one of the data x _i r (x j in a case where the other data x _j and the second query calculate. For example, it can be calculated using the following formula (1).

here,

Is a parameter.

Further, the similarity s (x _i , x _j ) is set as the reciprocal of the geometric mean, the reciprocal of the arithmetic mean, the reciprocal of the maximum value, etc. at r (x _i | x _j ) and r (x _j | x _i ). May be good.

Then, when the similarity calculation unit 130 finds the similarity for all the data pairs, it finds the similarity matrix that stores all the obtained similarity, and the similarity matrix and the data set.

To the rank score calculation unit 140.

The rank score calculation unit 140 calculates a rank score indicating the degree to which the data is related to the first query based on the similarity obtained by the similarity calculation unit 130 for each of the plurality of data to be searched. ..

Specifically, the rank score calculation unit 140 sets data for the first query q based on the similarity matrix acquired from the similarity calculation unit 130.

Calculate the rank score of all the data contained in.

The rank score can be calculated by using, for example, the method described in Non-Patent Document 3. Here, the similarity matrix

age,

And.

Further, when the similarity is obtained using the above equation (7), the elements of the similarity matrix are

And.

First, the rank score calculation unit 140 sets the diagonal component of A to 0.

Next, the rank score calculation unit 140 calculates the diagonal matrix D using the following equation (2).

Here, d _j is the sum of the degree of similarity with all the second queries when x _j is the search target as in the following equation (3).

The rank score calculation unit 140 uses the following equation (4) to perform

To calculate.

Next, the matrix w for storing the rank score is calculated according to the following equation (5). Here, the matrix w

And.

Here, I has an identity matrix,

Is a parameter, and for example, α = 0.99 can be adopted. Also,

Is also a parameter.

The rank score can also be calculated by using the iterative method described in Non-Patent Document 3. For example, the matrix w is calculated by repeating the calculation according to the following equation (6).

here,

Is. Also, the w to be calculated when the t-th iteration and w ^t, and ^w 0 = y. Further, for example, the number of iterations can be set to 30, and the matrix w calculated in the final iteration can be used as the calculation result by the rank score calculation unit 140.

The rank score can also be calculated by the method of solving the above equation (5) using the conjugate gradient method.

Then, the rank score calculation unit 140 searches for data in descending order of rank score based on the calculated matrix w.

Is sorted, and the sorting result is passed to the output unit 150.

The output unit 150 outputs the sorting result by the rank score calculation unit 140.

<Operation of the search device according to the first embodiment of the present invention>
FIG. 2 is a flowchart showing a search processing routine according to the first embodiment of the present invention.

When the first query is input to the input unit 100, the search device 10 executes the search processing routine shown in FIG.

First, in step S100, the input unit 100 accepts the input of the first query.

In step S110, the search unit 120 acquires a plurality of data to be searched from the database 110.

In step S120, the search unit 120 refers to each of the data included in the data set including the data which is the first query received in step S100 and the plurality of data to be searched obtained in step 100 as elements. , A rank, which is a rank related to the second query, is given to each of the data included in the data set when the data set is searched for as the second query.

In step S130, the similarity calculation unit 130 sets the similarity between one of the data and the other of the data for each pair of data included in the data set as the second query. It is calculated that the higher the rank, the larger the value given to the other data and the rank given to one of the data when the other data is used as the second query.

In step S140, the rank score calculation unit 140 indicates the degree to which the data is related to the first query based on the similarity obtained by the similarity calculation unit 130 for each of the plurality of data to be searched. Calculate the score.

In step S150, the rank score calculation unit 140 sorts the data to be searched in descending order of rank score based on the rank score obtained in step S140.

In step S160, the output unit 150 outputs the sorting result obtained in step S150.

As described above, according to the search device according to the present embodiment, each of the data pairs included in the data set containing the input first query data and the plurality of data to be searched as elements. The degree of similarity between one of the relevant data and the other of the relevant data was defined as the rank given to the other relevant data when one of the relevant data was used as the second query, and the other of the relevant data as the second query. By calculating so that the higher the rank, the larger the rank given to one of the relevant data in the case, the similarity between the related search target and the unrelated search target is suppressed to a low level. The rank score of irrelevant search targets can be calculated more accurately, the accuracy of re-ranking can be improved, and the accuracy of search can be improved.

In the present embodiment, the similarity calculation unit 130 uses the similarity between one of the data and the other as the second query for each of the pairs of data included in the data set. When both the rank given to the other relevant data and the rank given to one of the relevant data when the other relevant data is used as the second query are higher than the predetermined rank. It may be calculated that the higher the rank is, the larger the value is taken, and when at least one of them is lower than the predetermined rank, the smaller value is taken.

Specifically, the similarity calculation unit 130 assigns the similarity s (x _i , x _j ) to the rank r (x) given to the other data x _j when one data x _i is used as the second query. _i | x _j ) and the rank r (x _j | x _i ) given to one data x _i when the other data x _j is used as the second query are both higher ranks than the predetermined rank. In some cases, the higher the rank, the larger the value, and when both ranks are not higher than the predetermined rank, the smaller value is calculated. For example, it can be calculated using the following equation (7).

Here, k is a threshold value of a positive integer having a predetermined rank, for example, k = 30, and _sk (x _i , x _j ) is the degree of similarity when the equation (7) is used. Is shown.

When equation (7) is used, if both r (x _i | x _j ) and r (x _j | x _i ) are not higher than a predetermined rank, the similarity is calculated to take a small value. By doing so, the similarity between the related search target data and the unrelated search target data can be suppressed to be lower, and a more accurate similarity can be obtained.

<Overview of the search device according to the second embodiment of the present invention>
The outline of the second embodiment of the present invention will be described.

The accuracy of the search target reranking method strongly depends on the accuracy of the contribution of the top results to the reranking. When the degree of contribution to the reranking of the high-ranking false positive results is calculated to be high, the unrelated search target having a high degree of similarity to the high-ranking result is calculated to have a high rank score, which adversely affects the accuracy of the reranking. For example, the method described in Non-Patent Document 1 uniformly treats the degree of contribution to the reranking of all top results. In the method described in Non-Patent Document 2, a weighted value inversely proportional to the first rank of each higher-ranking result is calculated as the degree of contribution to the re-ranking of the higher-ranking result. However, the accuracy of the first rank strongly depends on the accuracy of the first search, and when the accuracy of the first search is low, the first rank is high and a high weighted value is given to the top result of false positive. Probability is high. Such high-ranked false-positive results of the first rank still have a high degree of contribution to re-ranking, and often adversely affect the accuracy of re-ranking.

In addition, the accuracy of the search target reranking method strongly depends on the accuracy of the similarity between the top result and the search target. When the degree of similarity between the high-ranking true positive result and the unrelated search target is calculated to be high, the rank score of the search target is calculated to be high, and the accuracy of reranking is lowered. For example, in the methods described in Non-Patent Documents 1 and 2, the similarity is calculated based on the rank given to the higher-ranked result or the search target in the second search.

However, if the accuracy of the second search is low, there is a problem that even if the search target is irrelevant, the similarity with the top result of true positive is calculated to be high, and the accuracy of reranking is likely to be adversely affected. ..

In the present embodiment, for each of the data included in the data set containing the input first query data and the plurality of data to be searched as elements, the data is used as the second query for the data set. Search and give a rank to all data.

Next, in the second data set in which the upper results in the first query are acquired from the search targets and the first query and all the upper results are combined, the same method as in the first embodiment is used to obtain all the data. Calculate the rank score (first rank score).

For each of the pairs in which the data included in the plurality of data to be searched and the data included in the second data set are combined, the similarity is determined by using the same method as in the first embodiment based on the rank. For each of the search targets, the weighted sum of the degree of similarity between the search target and all the data is calculated based on the first rank score of all the data in the second data set, and this is used as the second rank score.

With such a configuration, it is possible to suppress the contribution of false positive top results to the reranking and the similarity between the true positive top results and unrelated search targets to a low level, and improve the accuracy of reranking. ..

Further, the calculation of the first rank score, which has a large amount of calculation as compared with the configuration of the first embodiment and the method of the prior art (for example, the method of Non-Patent Document 3), is limited to only the higher-ranked results which are less than the number of data to be searched. Therefore, the amount of calculation can be suppressed to a small level, and the search speed can be increased.

<Structure of the search device according to the second embodiment of the present invention>
The configuration of the search device 20 according to the second embodiment of the present invention will be described. The same components as those of the search device 10 according to the first embodiment are designated by the same reference numerals, and detailed description thereof will be omitted.

The search device 20 is composed of a computer including a CPU, a RAM, and a ROM that stores a program for executing a search processing routine described later, and is functionally configured as shown below.

As shown in FIG. 3, the search device 20 according to the present embodiment includes an input unit 100, a database 110, a search unit 120, a first similarity calculation unit 230, a first rank score calculation unit 240, and a first rank score calculation unit 240. The two similarity calculation unit 250, the second rank score calculation unit 260, and the output unit 150 are provided.

The search unit 120 further includes the generated rank matrix and the generated data set.

Is passed to the second similarity calculation unit 250. Here, in the present embodiment, a data set

Is called the first data set.

The first similarity calculation unit 230 uses one of the relevant data for each of the data that is the first query and the data pair included in the second data set whose elements are the data to be searched with the top N ranks. The degree of similarity between and the other data is the rank given to the other data when one data is used as the second query, and the one data when the other data is used as the second query. The higher the rank, the larger the value given to.

Specifically, the first similarity calculation unit 230 first searches data when the first query q (x ₀ ) included in the rank matrix calculated by the search unit 120 is used.

Based on the rank r (x ₁ | x ₀ ), r (x ₂ | x ₀ ), ..., R (x _n | x ₀ ) given to, the top m pieces of the data to be searched in the first query. Get the data of. For example, when acquiring top 100 a (m = 100) pieces of data, _r | acquires _(x i x ₀₎ ≦ 100 and becomes _{x i.} Here, the obtained top results are

And.

Then, the first similarity calculation unit 230 includes the first query q and the acquired higher-level result.

A second data set with and as one set

To create.

Next, the first similarity calculation unit 230 uses the second data set.

For each data pair to be included in, the data pairs _{z i, _{z j}} for similarity _{s (z} i, _{z j),} and the other in the case where one of the data _{z i} is defined as a second query data _{z j} in granted rank _r and | | _{(z i z} j) and _(z i _{z j),} the other data _{z j} the second query and the rank given to one of the data _{z i} when _r Calculate so that the higher the rank of both, the larger the value.

For example, it calculated using the above equation (1), also the similarity s (z _{i, z j),} and the reciprocal of the geometric mean reciprocal of the arithmetic mean, or as the reciprocal of the maximum value and the like.

Further, the first similarity calculation unit 230 uses one of the data pairs included in the second data set whose elements are the data that is the first query and the data that is the search target of the top N ranks. The degree of similarity between the data and the other data is the rank given to the other data when one of the data is used as the second query, and one when the other data is used as the second query. When the ranks given to the relevant data are both higher than the predetermined rank, the higher the rank, the larger the value is calculated, and when both are higher than the predetermined rank, both are calculated. The higher the rank, the larger the value may be calculated, and if at least one of the ranks is lower than the predetermined rank, the smaller value may be calculated.

Specifically, the first similarity calculation unit 230 uses the second data set.

For each data pair to be included in, the data pairs _{z i, _{z j}} for similarity _{s (z} i, _{z j),} and the other in the case where one of the data _{z i} is defined as a second query data _{z j} in granted rank _r and | | _{(z i z} j) and _(z i _{z j),} the other data _{z j} the second query and the rank given to one of the data _{z i} when _r However, when both ranks are higher than the predetermined rank, the higher the rank, the higher the value is calculated, and when both are higher than the predetermined rank, the higher the rank, the higher the value. If at least one of them has a rank lower than a predetermined rank, it is calculated to take a smaller value.

For example, using the method of the above formula (7), the similarity _{s k (z i, z j} ) can be obtained.

Then, the first similarity calculation unit 230 obtains the first similarity matrix that stores all the obtained similarity, and the first similarity matrix and the second data set.

To the second rank score calculation unit 260.

Further, the first similarity calculation unit 230 uses the second data set.

Is passed to the second similarity calculation unit 250.

The first rank score calculation unit 240 indicates the degree to which the data is related to the first query based on the similarity obtained by the first similarity calculation unit 230 for each data in the second data set. Calculate the score.

Specifically, the first rank score calculation unit 240 is based on the first similarity matrix acquired from the first similarity calculation unit 230 by the same method as the rank score calculation unit 140 according to the first embodiment. Second data set for the first query q

The first rank score, which is the rank score for each of the data contained in, is calculated.

Then, the first rank score calculation unit 240 passes the first rank score matrix storing the first rank score to the second rank score calculation unit 260.

The second similarity calculation unit 250 uses one of the data and the other for each of the pairs of the data included in the plurality of data to be searched and the data included in the second data set. The second similarity, which is the degree of similarity with the data, is the rank given to the other data when one of the data is used as the second query, and one when the other data is used as the second query. It is calculated that the higher the rank given to the relevant data, the larger the value.

Specifically, the second similarity calculation unit 250 uses a plurality of data to be searched.

The data x _i contained in and the second data set

For each pair the combination of the data _{z j} included in, the data pairs _{z i, _{x j}} for the second degree of similarity _{s (x} i, _{z j),} and one of the data _{x i} second query the rank given to the other data z _j in the case where r (z i _{| x i)} and the other data z _j the second one in the case of the query data x _i to be applied rank r ( It is calculated so that the higher the rank of both x _i | z _j ), the larger the value.

For example, it calculated using the above equation (1), and the second similarity s (x _{i, z j),} and the reciprocal of the geometric mean reciprocal of the arithmetic mean, or as the reciprocal of the maximum value and the like.

Then, the second similarity calculation unit 250 obtains a second similarity matrix that stores all the obtained second similarity, and passes the second similarity matrix to the second rank score calculation unit 260.

The second rank score calculation unit 260 indicates the degree to which the data is related to the first query based on the first rank score of the data and the second similarity for each of the plurality of data to be searched. Calculate the second rank score.

Specifically, the second rank score calculation unit 260 sets the second data set for each of the data to be searched.

Based on the first rank score for each of the data contained in, the weighted sum of the second similarity between the data to be searched and the pair of data contained in the second data set is calculated, and the second rank score is calculated. And.

Here, the first rank score matrix

If represents the second rank score data _{x i} of the search target p _{(x i).}

For example, the second rank score is calculated using the following equations (8) and (9).

Then, the second rank score calculation unit 260 searches for data in descending order of the second rank score.

Is sorted, and the sorting result is passed to the output unit 150.

<Operation of the search device according to the second embodiment of the present invention>
FIG. 4 is a flowchart showing a search processing routine according to the second embodiment of the present invention. The same processing as the search processing routine according to the first embodiment is designated by the same reference numerals and detailed description thereof will be omitted.

In step S230, the first similarity calculation unit 230 uses the first query data and each of the data pairs included in the second data set whose elements are the data to be searched with the highest N ranks. The degree of similarity between one of the data and the other data is the rank given to the other data when one of the data is used as the second query, and the degree of similarity between the other data is used as the second query. It is calculated so that the higher the rank given to the data, the larger the value.

Further, in step S230, the first similarity calculation unit 230 is a pair of data included in the second data set whose elements are the data that is the first query and the data that is the search target of the top N ranks. For each, the similarity between the data on one side and the data on the other side is defined as the rank given to the data on the other side when one piece of data is used as the second query, and the data on the other side is referred to as the second query. If one of the ranks given to the relevant data in the case of the above is higher than the predetermined rank, the higher the rank, the larger the value is calculated, and both are higher than the predetermined rank. In some cases, the higher the rank, the larger the value, and if at least one of them is lower than the predetermined rank, the smaller the value may be calculated.

In step S240, the first rank score calculation unit 240 indicates the degree to which the data is related to the first query for each data in the second data set based on the similarity obtained in step S230. Calculate the score.

In step S242, the second similarity calculation unit 250 uses one of the relevant pairs of the data included in the plurality of data to be searched and the data included in the second data set. The second similarity, which is the similarity between the data and the other data, is the rank given to the other data when one of the data is used as the second query, and the other data is referred to as the second query. In this case, the higher the rank, the larger the value given to the data.

In step S244, the second rank score calculation unit 260 uses the first rank score of the data and the second similarity obtained in step S242 for each of the plurality of data to be searched. Calculate a second rank score that indicates how much the data is relevant to the first query.

In step S250, the second rank score calculation unit 260 sorts the data to be searched in descending order of the second rank score obtained in step S244.

As described above, according to the search device according to the present embodiment, each of the data included in the first data set having the input first query data and the plurality of data to be searched as elements. When a search is performed on the first data set using the data as the second query, each of the data included in the first data set is given a rank, which is the rank related to the second query, and the first For each of the pair of data included in the second data set whose elements are the data that is the first query and the data that is the search target of the top N ranks, the similarity calculation unit has one of the relevant data and the other of the data. The degree of similarity with the data is given to the rank given to the other data when one of the data is used as the second query and to one of the data when the other data is used as the second query. The higher the rank, the larger the value is calculated so that the first rank score calculation unit obtains the similarity obtained by the first similarity calculation unit for each data in the second data set. Based on this, a first rank score indicating the degree to which the data is related to the first query is calculated, and each of the data contained in the plurality of data to be searched and each of the data contained in the second data set are obtained. For each of the combined pairs, the second similarity, which is the similarity between one data and the other data, is the rank given to the other data when one data is used as the second query. , When the other relevant data is used as the second query, the higher the rank of the other relevant data, the larger the value is calculated, and each of the plurality of data to be searched is calculated. , By calculating a second rank score that indicates the degree to which the data is relevant to the first query, based on the first rank score of the data and the second similarity, with respect to the reranking of the top false positive results. It is possible to suppress the degree of contribution and the degree of similarity between the high-ranking true positive results and the unrelated search target to a low level, improve the accuracy of re-ranking, and improve the accuracy of the search.

Further, since the calculation of the first rank score having a large amount of calculation is limited to only the high-ranking results smaller than the number of data to be searched, the amount of calculation can be suppressed to be small and the search speed can be increased.

The present invention is not limited to the above-described embodiment, and various modifications and applications are possible without departing from the gist of the present invention.

For example, in the second embodiment, the second similarity calculation unit 250 is a pair of each of the data included in the plurality of data to be searched and each of the data included in the second data set. The second similarity, which is the similarity between the one data and the other data, is the rank given to the other data when one of the data is used as the second query, and the other data. Is used as the second query, and when one of the ranks given to the relevant data is higher than the predetermined rank, the higher the rank, the larger the value, and at least one of them is calculated to take a larger value. It may be calculated so as to take a small value when the rank is lower than the predetermined rank.

Specifically, the second similarity calculation unit 250 uses a plurality of data to be searched.

The data x _i contained in and the second data set

For each pair the combination of the data _{z j} included in, the data pairs _{z i, _{x j}} for the second degree of similarity _{s (x} i, _{z j),} and one of the data _{x i} second The rank r (z _j | x _i ) given to the other data z _j in the case of a query and the rank r (g) given to one data x _i in the case of using the other data z _j as a second query. It is calculated that x _i | z _j ) takes a larger value as the rank is higher than the predetermined rank when both are higher than the predetermined rank, and both are higher ranks when both are higher than the predetermined rank. The higher the value, the larger the value, and when at least one of them has a rank lower than the predetermined rank, the smaller value is calculated.

For example, using the method of the above formula (7), the second similarity _{s k (x i, z j} ) can be obtained.

Further, in the second embodiment, the first similarity calculation unit 230 performs a higher search.

And the second data set

However, the search unit 120 may generate the above and pass it to the first similarity calculation unit 230 and the second similarity calculation unit 250, respectively.

In addition, the second similarity calculation unit 250 also performs a high-level search.

And the second data set

It may be configured to create and.

Further, although described as an embodiment in which the program is pre-installed in the specification of the present application, it is also possible to provide the program by storing it in a computer-readable recording medium.

10 Search device 20 Search device 100 Input unit 110 Database 120 Search unit 130 Similarity calculation unit 140 Rank score calculation unit 150 Output unit 230 First similarity calculation unit 240 First rank score calculation unit 250 Second similarity calculation unit 260 2 rank score calculation department

Claims (8)

When a search is performed on the data set using the data as the second query for each of the data which is the input first query and the data included in the data set whose elements are a plurality of data to be searched. , A search unit that assigns a rank, which is a rank related to the second query, for each of the data included in the data set.
For each of the pairs of data contained in the data set, the similarity between the one data and the other data is the rank given to the other data when one of the data is used as the second query. , A similarity calculation unit that calculates so that the higher the rank, the larger the rank given to one of the data when the other data is used as the second query.
With respect to each of the plurality of data to be searched, a rank score calculation unit that calculates a rank score indicating the degree to which the data is related to the first query based on the similarity obtained by the similarity calculation unit. ,
Search device including. The similarity calculation unit
For each of the pairs of data contained in the data set, the similarity between the one data and the other data is the rank given to the other data when one of the data is used as the second query. , When the other data is used as the second query and the rank given to one of the data is higher than the predetermined rank, the higher the rank, the larger the value. The search device according to claim 1, wherein the search device is calculated so as to take a smaller value when at least one of them has a rank lower than the predetermined rank. For each of the data which is the input first query and the data included in the first data set whose elements are a plurality of data to be searched, the search for the first data set is performed using the data as the second query. A search unit that assigns a rank, which is a rank related to the second query, for each of the data included in the first data set when the data is performed.
For each of the data that is the first query and the data pair included in the second data set whose elements are the data to be searched with the top N ranks, one of the data and the other of the data The degree of similarity is the rank given to the other data when one of the data is used as the second query, and the rank given to one of the data when the other data is used as the second query. However, the first similarity calculation unit that calculates so that the higher the rank, the larger the value,
For each data in the second data set, a first rank that calculates a first rank score indicating the degree to which the data is related to the first query is calculated based on the similarity obtained by the first similarity calculation unit. Score calculation department and
For each of the pairs in which each of the data included in the plurality of data to be searched and each of the data included in the second data set are combined, the degree of similarity between the one data and the other data A certain second similarity was given to the rank given to the other data when one of the data was used as the second query, and to one of the data when the other data was used as the second query. The second similarity calculation unit that calculates so that the higher the rank, the larger the value,
For each of the plurality of data to be searched, a second rank score indicating the degree to which the data is related to the first query is calculated based on the first rank score of the data and the second similarity. 2nd rank score calculation unit and
Search device including. The first similarity calculation unit is
For each of the pairs of data contained in the second data set, the similarity between the one data and the other data was given to the other data when one data was used as the second query. When the rank and the rank given to one of the data when the other data is used as the second query are both higher than the predetermined rank, the higher the rank, the larger the value. Calculate so that if both ranks are higher than the predetermined rank, the higher the rank, the higher the value, and if at least one of them is lower than the predetermined rank, the smaller the value. Calculate to
The second similarity calculation unit is
For each of the pairs in which each of the data included in the plurality of data to be searched and each of the data included in the second data set are combined, the degree of similarity between the one data and the other data A certain second similarity was given to the rank given to the other data when one of the data was used as the second query, and to one of the data when the other data was used as the second query. When both ranks are higher than the predetermined rank, the higher the rank, the larger the value is calculated, and when at least one of them is lower than the predetermined rank, the smaller value is taken. The search device according to claim 3, wherein the search device is calculated as follows. The search unit searches for the data set as the second query for each of the data included in the data set including the input first query data and the plurality of data to be searched. A search step for assigning a rank, which is a rank related to the second query, for each of the data included in the data set when the data is performed.
For each of the pairs of data included in the data set, the similarity calculation unit determines the similarity between the one data and the other data, and the other data when one of the data is used as a second query. The similarity calculation step is calculated so that the higher the rank, the larger the rank given to the data and the rank given to one of the data when the other data is used as the second query. ,
The rank score calculation unit calculates a rank score indicating the degree to which the data is related to the first query based on the similarity obtained by the similarity calculation unit for each of the plurality of data to be searched. Rank score calculation steps to be performed and
Search methods that include. The similarity calculation step is
For each of the pairs of data contained in the data set, the similarity between the one data and the other data is the rank given to the other data when one of the data is used as the second query. , When the other data is used as the second query and the rank given to one of the data is higher than the predetermined rank, the higher the rank, the larger the value. The search method according to claim 5, wherein the search method is calculated so as to take a smaller value when at least one of the ranks is lower than the predetermined rank. The search unit uses the data as the second query for each of the data included in the first data set whose elements are the input first query data and the plurality of data to be searched, and the first data. A search step for assigning a rank, which is a rank related to the second query, for each of the data included in the first data set when a search is performed on the set, and
One of the first similarity calculation units is for each of the data that is the first query and the data pair that is included in the second data set that includes the data to be searched for the top N ranks. The degree of similarity between the data and the other data is the rank given to the other data when one of the data is used as the second query, and one of the ranks given to the other data when the other data is used as the second query. The first similarity calculation step, which calculates so that the higher the rank of the rank assigned to the data, the larger the value.
The first rank score calculation unit indicates the degree to which the data is related to the first query based on the similarity obtained by the first similarity calculation unit for each data of the second data set. The first rank score calculation step to calculate the rank score and
For each of the pairs of data included in the plurality of data to be searched and each of the data included in the second data set, the second similarity calculation unit has one of the data and the other. The second similarity, which is the similarity with the data, is the rank given to the other data when one of the data is used as the second query, and the second similarity when the other data is used as the second query. A second similarity calculation step for calculating that the higher the rank of the rank assigned to the data, the larger the value.
The degree to which the data is related to the first query based on the first rank score of the data and the second similarity for each of the plurality of data to be searched by the second rank score calculation unit. The second rank score calculation step to calculate the second rank score indicating
Search methods that include. A program for causing a computer to function as each part of the search device according to any one of claims 1 to 4.