JP5432936B2 - Document search apparatus having ranking model selection function, document search method having ranking model selection function, and document search program having ranking model selection function - Google Patents

Description

  The present invention relates to an apparatus and a method for presenting document search results.

  In a search system such as a web search system, there are many factors (score factors such as a score based on a query frequency such as TF-IDF (Term Frequency-Inverse Document Frequency) and a score based on link analysis such as PageRank). The search score used for the final ranking is calculated (see Non-Patent Document 1). A method of presenting the search results obtained by ranking by arranging the calculated search scores in descending order is widely used.

  Since the optimal ranking model is different for each query, there is a method of selecting and applying the optimal ranking model for each query. In Non-Patent Document 2, a ranking model is generated for each query, and a ranking model with the shortest Euclidean distance is selected for the input query in the feature expression of the query. At this time, for example, the technique of Non-Patent Document 3 is used as each ranking model generation method.

  Note that the conversion matrix used in the document search apparatus of the present invention is described in Non-Patent Document 4 below.

Hiroshi Takeno, Takashi Inoue, “Distributed high-speed information collection / full-text search system InfoBee / Evangelist”, NTT R & D Vol. 52 no. 2 2003, pp. 78≡84. Xiubo Geng, Tie-Yan Liu, Tao Qin, Andrew Arnold, Hang Li and Heung-Yeung Shum, "Query Dependent Ranking Using K-Nearest Neighbor", In Proceedings of the 31st annual international ACM SIGIR conference on Research and development in information retrieval (SIGIR '08), 2008, pp. 115-122. Thorsten Joachims, “Optimizing Search Engineers using Clickthrough Data”, In Proceedings of the height of the ACM International Conference on Knowledge. 133-142. Eric P. Xing and Andrew Y. Ng and Michael I. Jordan and Stuart Russell, “Distance Metric Learning, with Application to Clustering with Side-Information. Proceedings of the 16th Announcement”. 505-512.

  The query similarity calculation method used in the prior art to select a ranking model for each query is the distance in the feature space set in advance, and the nearest neighbor in this feature space for the input query Is based on the assumption that the ranking model generated using

  However, since the nearest model in the query feature expression and the actually optimal model do not always match, the similarity cannot be calculated appropriately, and the optimal model that achieves high-precision ranking There is a problem that the model cannot be selected.

  The present invention solves the above-described problems, and an object thereof is to provide a document search apparatus, method, and program having a ranking model selection function capable of selecting an optimal ranking model for an input query. .

  In order to solve the above problems, a document retrieval apparatus having a ranking model selection function according to the present invention includes a query expression database storing query expression data expressed by M-dimensional features for each of N queries, and N A training data database storing training data having a matching degree of a document search result with respect to a query and a feature expression represented by an M-dimensional vector, and the training data as an input, the feature expression of each query A ranking function generating means for generating a ranking model holding weights for the ranking model database, and the training data and ranking model as inputs, and the maximum search evaluation index value for all models in the ranking model database The optimal ranking model during training showing the highest accuracy in training data A training-time optimal model selection means for generating and building a training-time optimal model database having data of a pair of the training-time optimal ranking model and the training data query; and the query expression database and the training-time optimal model database. And learning and generating a transformation matrix that minimizes the distance between the query in the query expression database and the query of the optimal ranking model for training in the optimal model database for training corresponding to the query. A distance learning means for constructing a transformation matrix database, and input each data of the query expression database, ranking model database and transformation matrix database, and for each query in the query expression database, between the queries using the transformation matrix Calculate the similarity of and have the maximum similarity An optimal model database creating means for selecting an Eri, acquiring a ranking model of the selected query from the ranking model database, and constructing an optimal model database using the acquired ranking model as an optimal model for the query, and a Web page in advance A document index database storing a document index created based on a document collected from the database, a search result set for the input search query is acquired from the document index database, and the search result set and a plurality of score factors A query processing means for calculating a score factor value matrix, a score factor value matrix calculated by the query processing means, and each data of the ranking model database and the optimal model database, and corresponding to the input search query Optimal model to do The search score vector is calculated by accumulating the weight as the ranking model in the ranking model database corresponding to the query of the acquired optimum model and the score factor value matrix. Search score calculation means and search result presentation means for presenting search results for the input query in descending order of the search score calculated by the search score calculation means.

  According to the present invention, since the query feature space is converted so that the query that maximizes the search evaluation index is close to the neighborhood in the feature space, the similarity calculation with respect to the input query is improved. The performance of model selection can be improved and the accuracy of search ranking can be improved.

1 is a configuration diagram of an entire document search apparatus according to an embodiment of the present invention. The block diagram of the apparatus which produces the optimal model DB of FIG. The block diagram of the apparatus which produces | generates transformation matrix DB of FIG. The flowchart which shows the flow of a process of the optimal model selection part 120 at the time of training of FIG. The flowchart which shows the flow of a process of the optimal model DB creation part 140 of FIG. 3 is a flowchart showing a flow of processing of the document search apparatus in FIG. 1.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings, but the present invention is not limited to the following embodiments. First, an overview of the overall configuration of an embodiment of the present invention will be described with reference to FIGS.

  As shown in FIG. 1, the document search apparatus 100 according to the present embodiment includes a document index DB (database) 107 that stores document index data created based on documents collected in advance from Web pages, ranking model data, and the like. Is stored in the ranking model DB 103, the optimal model DB 106 storing the data of the optimal model, the query processing unit 150 as the query processing unit, the search score calculation unit 160 as the search score calculation unit, and the search result as the search result presentation unit A presentation unit 170 is provided.

  The optimum model DB 106 in FIG. 1 is processed by the optimum model DB creation unit 140 as the optimum model DB creation unit based on the data stored in the query expression DB 101, the ranking model DB 103, and the transformation matrix DB 105 shown in FIG. Built.

  The ranking model DB 103 in FIG. 1 is constructed by the processing of the ranking function generation unit 110 serving as a ranking function generation unit based on data stored in the training data DB 102 illustrated in FIG.

  The transformation matrix DB 105 in FIG. 2 is based on the data stored in the query expression DB 101, the training data DB 102, the ranking model DB 103, and the training optimum model DB 104 shown in FIG. It is constructed by the respective processes of the optimum model selection unit 120 during training as means and the distance learning unit 130 as distance learning means.

  Note that the transformation matrix generation device 115 is configured by the ranking model DB 103, the training optimal model DB 104, the ranking function generation unit 110, the training optimal model selection unit 120, and the distance learning unit 130 of FIG.

  The document search apparatus 100 shown in FIGS. 1 to 3 is configured by a computer, for example, and is a normal computer hardware resource, for example, ROM, RAM, CPU, input device, output device, display device, communication interface, hard disk, recording medium And a driving device thereof.

  As a result of the cooperation between the hardware resource and the software resource (OS, application, etc.), as shown in FIGS. 1 to 3, the document search apparatus 100 performs a query expression DB 101, a training data DB 102, a ranking model DB 103, and a training time. Optimal model DB 104, transformation matrix DB 105, optimal model DB 106, document index DB 107, ranking function generation unit 110, optimal model selection unit 120 during training, distance learning unit 130, optimal model DB creation unit 140, query processing unit 150, search score calculation Unit 160 and search result presentation unit 170 are implemented.

  The query expression DB 101, the training data DB 102, the ranking model DB 103, the training optimum model DB 104, the transformation matrix DB 105, the optimum model DB 106, and the document index DB 107 are constructed in storage means / storage means such as a hard disk or RAM. .

  Next, details of the apparatus configured as described above will be described.

  First, in FIG. 3, the transformation matrix generation device 115 receives the query expression DB 101 and the training data DB 102 as inputs, and generates the transformation matrix DB 105. An example of the data structure of the training data DB 102 is shown in Table 1.

  In this example, training data to which a fitness level is assigned to each of search results (documents) for N queries is shown. Each line represents the degree of conformity and feature expression of a document for the query. The document is represented by M features in this example and can be represented using M-dimensional vectors.

<Ranking function generator 110>
The ranking function generation unit 110 receives the training data DB 102 as an input, and generates a ranking model DB 103. For the ranking function generation unit 110, for example, the technique of Non-Patent Document 3 can be used. An example of the data structure of the ranking model DB 103 is shown in Table 2.

  In Table 2, the i-th line represents the ranking model obtained by inputting the training data of the query qi to the ranking function generation unit 110. Table 2 is an example of a ranking model for each of the N queries. The ranking model can be expressed as a weight for the feature expression of the input document. That is, in the example of Table 2, wi represents the weight for xi of Table 1. When a document is represented by an M-dimensional feature expression x, the ranking model can be similarly represented by an M-dimensional weight w.

<Optimum model selection unit 120 during training>
The training optimal model selection unit 120 receives the training data DB 102 and the ranking model DB 103 as inputs, performs the processing shown in steps S121 to S124 in FIG. 4 and outputs the training optimal model DB 104.

  Table 3 shows an example of the data structure used internally by the training optimal model selection unit 120, and Table 4 shows an example of the data structure of the training optimal model DB 104 output by the training optimal model selection unit 120.

  First, in step S121 of FIG. 4, an unprocessed query q is acquired from the training data DB 102.

  Next, in step S122, a query p that gives the maximum evaluation value (search evaluation index value) is determined for all the ranking models included in the ranking model DB 103.

  Next, in step S123, the query p (for example, the query ID of the optimal model in Table 4) and q (for example, the query ID of Table 4) are output to the optimal model DB104 during training.

  In step 124, the processes in steps S121 to S123 are repeatedly executed until all the queries in the training data DB 102 are processed.

  For example, the training data of the highest-order query having a large search evaluation index excluding the query can be used for each query. Here, as the search evaluation index, MAP (Mean Average Precision), NDCG (Normalized Disclosed Cumulative Gain), or the like can be used. Hereinafter, for explanation, MAP is used as a search evaluation index.

By the processing as described above, a ranking model result indicating the highest accuracy in the training data (training data DB 102) is generated. In the example of Table 3, for query q ₁ , a MAP value of 0.7 by the ranking model generated using query q ₂ and a MAP value of 0.4 by the ranking model generated using query q _N Indicates that is obtained.

<Distance learning unit 130>
The distance learning unit 130 receives the query expression DB 101 and the training optimal model DB 104 as inputs, and outputs a transformation matrix DB 105. A data structure example of the query expression DB 101 is shown in Table 5, and a data structure example of the transformation matrix DB 105 is shown in Table 6.

  In the example of Table 5, the query expression DB 101 stores feature expressions of N queries, and the value of fi in a certain query indicates the value of the i-th feature of the query. In this example, each query is represented by M-dimensional features.

  When the query expression vector of the query qi is xi and the query expression vector of the query qj is xj, the distance between the two query expression vectors is

  Can be used to calculate. Here, the transformation matrix A is an M-dimensional square matrix that maps the M-dimensional feature space to the M-dimensional feature space. When A = I, the transformation matrix A is the Euclidean distance.

Table The third example, for q _1, because the ranking model generated using q ₂ showed the highest MAP value, the query expression vectors x ₁ of q _1, query representation vector x ₂ of q ₂ The conversion matrix A is learned so as to reduce the distance of. In this way, a ranking model showing the best result for all the queries qi (i = 1... N) is selected, and a transformation matrix is generated so that the distance between the selected query and the query is minimized. For example, the technique of Non-Patent Document 4 can be used to generate the transformation matrix A.

  Next, details of the optimum model DB creation apparatus in FIG. 2 will be described.

<Optimum model DB creation unit 140>
The optimum model DB creation unit 140 receives the transformation matrix DB 105 generated by the transformation matrix generation device 115 of FIG. 3, the query expression DB 101, and the ranking model DB 103, respectively, and performs the processes shown in steps S141 to S146 of FIG. The optimal model DB 106 is output.

  An example of the data structure of the optimal model DB 106 is shown in Table 7.

Table 7 shows an example in which the optimal model is output for the number of queries q _{N + L} that is larger than the number of queries N in the optimal model DB 104 during training. By constructing q _{N + L} optimal models in this way, it is possible to cope with a case where the number of input queries during document search processing described later is greater than q _N.

  First, in step S141 in FIG. 5, an unprocessed query q is acquired from the query expression DB 101.

  In step S142, the similarity d is calculated using the transformation matrix DB 105 for each query included in the query expression DB 101.

  Next, in step S143, the query p having the maximum similarity d is selected.

  Next, in step S144, the ranking model w corresponding to the selected query p is acquired from the ranking model DB 103.

  In step S145, the optimal model for the query q is output to the optimal model DB 106 as the acquired w.

  In step S146, the processes in steps S141 to S146 are repeatedly executed until the processing of all the queries in the query expression DB 101 is completed.

  Next, details of the document search apparatus 100 of FIG. 1 will be described with reference to the flowchart of FIG.

<Query processing unit 150>
The query processing unit 150 receives a search query as input, acquires a search result set (document) including the search query from the document index DB 107, and calculates a score factor value matrix using the search result set and a plurality of score factors. (Step S150).

  Specifically, when N search result sets are acquired from the document index DB 107 using M score factors, the score factor value matrix is:

It expresses. Here, the i-th row of D represents the score factor value of the i-th search result. For example, d ₂₃ is the third score factor value for the second document.

<Search score calculation unit 160>
The search score calculation unit 160 receives the score factor value matrix D output from the query processing unit 150, the data of the ranking model DB 103, the data of the optimal model DB 106, and the input search query q _input as inputs.

The search score calculation unit 160 obtains the query ID q _best of the optimal model corresponding to the _input search query q _input from the optimal model DB 106, and calculates the weight w (score factor weight) of the query of the optimal model from the ranking model DB 103. The search score vector is obtained based on the score factor weight and the score factor value matrix D (step S160).

The search score vector s for use in the search ranking is obtained by the product of the score factor value matrix D and the score factor weight w ^(qbest) .

  That is, the search score si for the i-th document is

  Calculated by

<Search result presentation unit 170>
The search result presentation unit 170 receives the search score vector s, and presents the search result for the query in the descending order of the search score si (displays or outputs it as data) (step S170).

  In addition, some or all of the functions of each means in the document search apparatus having the ranking model selection function of the present embodiment are configured by a computer program, and the program is executed using the computer to realize the present invention. Of course, it is possible to configure the procedure in the document search method having the ranking model selection function of the present embodiment by a computer program and cause the computer to execute the program. The computer-readable recording medium, for example, FD (Floppy (registered trademark) Disk), MO (Magneto-Optical disk), ROM (Read Only Memory), memory card, CD (Compact D) isk) -ROM, DVD (Digital Versatile Disk) -ROM, CD-R, CD-RW, HDD, removable disk, etc., and can be stored or distributed. It is also possible to provide the above program through a network such as the Internet or electronic mail.

DESCRIPTION OF SYMBOLS 100 ... Document search apparatus 101 ... Query expression DB
102 ... Training data DB
103 ... Ranking model DB
104 ... Optimum model DB for training
105 ... Transformation matrix DB
106 ... Optimal model DB
107 ... Document index DB
DESCRIPTION OF SYMBOLS 110 ... Ranking function production | generation part 120 ... Optimum model selection part 130 at the time of training 130 ... Distance learning part 140 ... Optimal model DB creation part 150 ... Query processing part 160 ... Search score calculation part 170 ... Search result presentation part

Claims (3)

A query expression database storing query expression data expressed by M-dimensional features for each of the N queries;
A training data database in which training data having a matching degree of a document search result with respect to N queries and a feature expression represented by an M-dimensional vector are stored;
Ranking function generating means for generating a ranking model database by generating a ranking model having the training data as an input and holding a weight for the feature expression of each query;
The training data and the ranking model are input, the maximum search evaluation index value is given to all models in the ranking model database, and the training-time optimal ranking model indicating the highest accuracy in the training data is generated. A training optimal model selection means for constructing a training optimal model database having data of a pair of an optimal ranking model query and the training data query;
Each data of the query expression database and the optimal model database for training is input, and the distance between the query in the query expression database and the query of the optimal ranking model for training in the optimal model database for training corresponding to the query is minimum. Distance learning means that learns and generates a transformation matrix and constructs a transformation matrix database;
Using each data of the query expression database, ranking model database, and transformation matrix database as input, for each query in the query expression database, the similarity between the queries is calculated using the transformation matrix, and the maximum similarity is obtained. An optimal model database creating means for acquiring a ranking model of the selected query from the ranking model database and constructing an optimal model database using the acquired ranking model as an optimal model for the query;
A document index database in which a document index created based on a document collected in advance from a Web page is stored;
Query processing means for acquiring a search result set for the input search query from the document index database, and calculating a score factor value matrix from the search result set and a plurality of score factors;
The score factor value matrix calculated by the query processing means and each data of the ranking model database and the optimal model database are input, and an optimal model corresponding to the input search query is acquired from the optimal model database, A search score calculating means for calculating a search score vector by integrating a weight as a ranking model in the ranking model database corresponding to the obtained query of the optimal model and the score factor value matrix;
Search result presenting means for presenting search results for the input query in descending order of the search score calculated by the search score calculating means;
A document retrieval apparatus having a ranking model selection function. The ranking function generation means receives training data in a training data database in which training data having a fitness of a document search result for N queries and a feature expression represented by an M-dimensional vector are stored. A ranking function generation step of generating a ranking model holding weights for the feature expression of each query and constructing a ranking model database;
The training optimal model selection means takes the training data and ranking model as input, gives the maximum search evaluation index value for all models in the ranking model database, and shows the highest accuracy in training data indicating the highest accuracy in training data Generating a model and constructing a training optimal model database having data of a pair of the training optimal ranking model query and the training data query; and a training optimal model selection step;
The distance learning means inputs the query expression database storing query expression data expressed by M-dimensional features for each of the N queries and each data of the optimal model database at the time of training, and queries in the query expression database Learning and generating a transformation matrix that minimizes the distance from the query of the optimal ranking model for training in the optimal model database for training corresponding to the query, and a distance learning step of constructing the transformation matrix database;
Optimal model database creation means inputs each data of the query expression database, ranking model database, and transformation matrix database, and calculates the similarity between each query using the transformation matrix for each query in the query expression database. An optimal model that selects a query having the maximum similarity, acquires a ranking model of the selected query from the ranking model database, and constructs an optimal model database using the acquired ranking model as an optimal model for the query A database creation step;
The query processing means acquires a search result set for the input search query from a document index database in which a document index created based on a document previously collected from a Web page is stored, and the search result set and a plurality of scores A query processing step for calculating a score factor value matrix with factors,
The search score calculation means inputs the score factor value matrix calculated by the query processing means and each data of the ranking model database and the optimal model database, and selects the optimal model corresponding to the input search query as the optimal A search score calculation step of calculating a search score vector by accumulating the weight as a ranking model in the ranking model database corresponding to the acquired query of the optimal model and the score factor value matrix, obtained from the model database When,
A search result presenting step for presenting a search result for the input query in descending order of the search score calculated by the search score calculating unit;
A document retrieval method having a ranking model selection function characterized by comprising: A document search program having a ranking model selection function for causing a computer to function as each means according to claim 1.

Images