JP2012008899A - Retrieval query recommendation method, retrieval query recommendation device and retrieval query recommendation program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a recommendation query following the retrieval intention of a user.SOLUTION: This retrieval query recommendation device is provided with a determination part 11 for determining the change type of a retrieval behavior based on time-sequentially arranged retrieval queries, and a retrieval type determination part 12 is configured to determine a current change type by inputting the history of the retrieval queries, and a recommendation query creation part 13 is configured to create recommendation query candidates with respect to the nearest retrieval query in the history of the retrieval queries, and to determine the change type when the nearest retrieval query is changed to each recommendation query candidate, and to apply it to each recommendation query candidate, and a recommendation query selection part 14 is configured to hold a transition table showing the transition distribution of the change type, and to acquire the distribution rate of the change type whose transition is performed from the current change type, and to, when the prescribed number of recommendation queries are selected from the recommendation query candidates, select the recommendation query so that the distribution rate of the change type applied to the selected recommendation queries can be close to the distribution rate acquired from the transition table.

Description

  The present invention relates to a technique for recommending a search query.

  A number of search engine services having a search function based on a query are provided as a system for searching a document suitable for a purpose from a large amount of documents existing on the Internet. In these search engine services, based on a group of words given as a query by a user, a document having contents suitable for them is searched for and presented to the user. At this time, since the search intention of the user is transmitted to the system only by the query, the user must input a query suitable for the processing of the search engine while expressing his / her search intention when searching. However, since creating such an appropriate query is a task that requires a certain amount of knowledge and experience, there is a need for a system that recommends queries to users in order to support users who are unfamiliar with search engines. Exists.

  Many methods for recommending queries have been proposed so far. For example, in Patent Document 1, a site where access is concentrated in a click log related to a certain query is specified, and the titles and summary sentences of those sites are analyzed to extract a phrase that expands the query.

  Further, in Non-Patent Document 1, by analyzing a graph structure composed of a set of a query obtained by analyzing a click log and a URL of a document included in a search result, paying attention only to a peripheral structure of a query. Then, a phrase related to the query is acquired as a recommendation query.

JP 2010-055164 A

Imai and 5 others, “Ambiguous Query Recommendation Method for Web Search Services”, DEIM Forum 2010, IEICE, 2010, A1-2

  In the above conventional technology, only the query entered by the user is focused, so it is related to the user's search tendency, whether the user wants to obtain a more refined search result or a related search result. There was a problem that the recommended query was decided.

  The present invention has been made in view of the above, and an object of the present invention is to provide a recommendation query that more closely matches the user's search intention.

  The search query recommendation method according to the first aspect of the present invention includes a step of inputting a search query history of a user and determining a change type that classifies a change intention of the search query derived from a relationship between search queries, Creating a plurality of recommended query candidates for the latest search query in the search query history, and determining a change type when the search query is changed from the last search query to each of the plurality of recommended query candidates. A step of assigning the change type to each of the plurality of recommended query candidates, and a change type that transitions from the change type determined from the history of the search query with reference to a transition table that describes the distribution of the change type transition The distribution ratio of the change type assigned to the recommended query to be selected is obtained in the transition table. So close to the distribution ratio obtained from Le, and having the steps of: selecting a recommended query from the recommended query candidates.

  The search query recommendation device according to the second aspect of the present invention is a determination means for inputting a search query arranged in time series and determining a change type that classifies the change intention of the search query derived from the relationship between the search queries. A search type determination unit that inputs a search query history of the user and determines a current change type using the determination unit; and a plurality of recommended queries for the latest search query that is the latest in the search query history A candidate is created, a change type is determined when the search query is changed from the last search query to each of the plurality of recommended query candidates using the determination unit, and the change type is determined for each of the plurality of recommended query candidates. The change query that transitions from the current change type with reference to the transition table that describes the recommended query creation means to be assigned and the distribution of change type transitions Recommended query selection means for selecting a recommended query from the recommended query candidates so that the distribution ratio of the change type given to the recommended query to be selected is close to the distribution ratio acquired from the transition table. It is characterized by having.

  A search query recommendation program according to a third aspect of the present invention causes a computer to execute the search query recommendation method.

  According to the present invention, it is possible to provide a recommendation query more in line with a user's search intention.

It is a functional block diagram which shows the structure of the search query recommendation apparatus in this Embodiment. It is a functional block diagram which shows the structure of the search query recommendation apparatus of Example 1. FIG. It is a figure which shows the example of the search query log | history stored in a search query log database. It is a figure which shows the example of the total table which totaled the change of the change type of a search query. It is a flowchart which shows the flow of a process of a query change total part. It is a figure which shows the example of recommendation query candidate information. It is a figure which shows the example of a transition table. It is a flowchart which shows the flow of the process which produces a recommendation query. It is a functional block diagram which shows the structure of the search query recommendation apparatus of Example 2. FIG. An example of correct query change information is shown. It is a flowchart which shows the flow of the process which learns a change type determination model.

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

  FIG. 1 is a functional block diagram showing the configuration of the search query recommendation device in the present embodiment.

  The search query recommendation device 1 illustrated in FIG. 1 includes a determination unit 11, a search type determination unit 12, a recommended query creation unit 13, and a recommended query selection unit 14. Each unit included in the search query recommendation device 1 may be configured by a computer including an arithmetic processing device, a storage device, and the like, and the processing of each unit may be executed by a program. This program is stored in a storage device included in the search query recommendation device 1, and can be recorded on a recording medium such as a magnetic disk, an optical disk, or a semiconductor memory, or provided through a network.

  The determination unit 11 inputs a search query arranged in time series, and determines a change type of search behavior. The search action change type is a classification of the user's intention to change the search query, which is derived from the relationship between the search queries. For example, the search query is changed from “sports” to “baseball”. “Narrow down” intended to obtain more refined information, “Generalization” intended to obtain broader information, such as changing the search query from “Baseball” to “Sports”, and “Baseball” "Relation" intended to obtain relevant information, such as changing from "soccer" to "soccer", "switching" not applicable to any of the above.

  The search type determination unit 12 inputs the user's search query history and uses the determination unit 11 to acquire the change type of the user's search behavior.

  The recommended query creation unit 13 creates a plurality of recommended query candidates for the most recent search query in the user search query history, and determines the change type when the most recent search query is changed to each recommended query candidate. The change type is assigned to each recommended query candidate.

  The recommended query selection unit 14 holds a transition table indicating a transition distribution from a certain change type to the next change type, and inputs the current change type of the user from the search type determination unit 12 and starts from the current change type. The distribution of the change type distribution ratio obtained from the transition table when the distribution ratio of the change type to be transitioned is acquired from the transition table and a predetermined number of recommended queries are selected from the recommended query candidates. Select a recommendation query to be close to the ratio.

[Example 1]
FIG. 2 is a functional block diagram illustrating the configuration of the search query recommendation device according to the first embodiment.

  The search query recommendation device 2 illustrated in FIG. 2 includes a determination unit 21, a search type determination unit 22, a recommended query creation unit 23, a recommended query selection unit 24, and a query change aggregation unit 25. A search query log database 26 is connected to the query change totaling unit 25. In addition, each part with which the search query recommendation apparatus 2 is provided is good also as what is comprised by the computer provided with the arithmetic processing unit, the memory | storage device, etc., and the process of each part is performed by a program. This program is stored in a storage device included in the search query recommendation device 2, and can be recorded on a recording medium such as a magnetic disk, an optical disk, or a semiconductor memory, or provided through a network.

  First, the creation of a transition table used when selecting a recommended query will be described. The transition table is created by acquiring three consecutive search queries from the search query log database 26 and totaling the transitions of the search query change types.

  The search query log database 26 stores a search query input to a commercial web search engine or an in-site search function installed on the website, together with input time information input the search query and input user ID information. Yes. The user ID information uses information acquired using cookie information of a web browser or the like. Since the user ID information only needs to be able to determine the context of the search query, session information recorded on the web server side may be used. FIG. 3 shows an example of information stored in the search query log database 26.

  The query change totaling unit 25 accesses the search query log database 26, totals the transition distribution of the change type of the search query, and transmits a totaling table as a totaling result to the recommended query creating unit 23. Specifically, the query change totaling unit 25 extracts three search queries arranged in chronological order input by the same user from the search query log database 26, and changes from the first search query to the second search query. Each change type from the search query to the third search query is obtained, and transitions of the change types are aggregated. FIG. 4 shows an example of the aggregation table. The query change tabulation unit 25 uses the change type from the first search query to the second search query as the preceding change type, the change type from the second search query to the third search query as the subsequent change type, and the tabulation table. Add 1 to the value in the appropriate column above.

  Here, the flow of processing of the query change totaling unit 25 will be described. FIG. 5 is a flowchart showing a processing flow of the query change totaling unit 25.

  First, the query change totaling unit 25 acquires all user ID information included in the search query log database 26, removes duplicate user ID information, and extracts a list of user ID information (step S101). The extracted user ID information is stored in a buffer provided in the query change totaling unit 25.

  Subsequently, one piece of user ID information is extracted from the buffer, and search queries stored in pairs with the user ID information are extracted from the search query log database 26 in ascending order of input time information (step S102).

  The three search queries are extracted in order from the head of the set of obtained search queries, the type of advance change from the first search query to the second search query, and the succession from the second search query to the third search query. The change type is determined, and one value in the aggregation table corresponding to the combination of the preceding change type and the subsequent change type is added (step S103). Then, it is confirmed whether or not all search queries have been processed. If there are any search queries that have not been processed, the process returns to step S103 and continues (step S104).

  For example, if the set of search queries arranged in chronological order corresponding to a certain user ID information is “Tokyo Station, Tokyo Station Restaurant, Tokyo Station French, Tokyo Station Last Train, Taxi Telephone”, first, “Tokyo Station” A set of three search queries “Tokyo Station Restaurant” and “Tokyo Station French” are extracted. The change type of the first search query “Tokyo Station”, the second search query “Tokyo Station Restaurant”, the second search query “Tokyo Station Restaurant” and the third search query “Tokyo Station French” is determined. . Since the preceding change type from the first search query to the second search query has a word added, “narrow down” and the subsequent change type from the second search query to the third search query are one of the words. Since the part has been changed, it is determined as “related”. As a result, 1 is added to the value of the location where the preceding change type is “squeezed” and the subsequent change type is “related” in the aggregation table. After that, the “Tokyo Station Restaurant”, “Tokyo Station French”, “Tokyo Station Last Train” group, and “Tokyo Station French”, “Tokyo Station Last Train”, “Taxi Telephone” group are also processed.

  Then, it is confirmed whether or not the user ID information remains in the buffer. If it remains, the process returns to step S102 to continue the processing (step S105).

  The aggregation table created by the query change aggregation unit 25 is transmitted to the recommended query selection unit 24 and converted into a transition table by the recommended query selection unit 24.

  Next, creation of a recommendation query will be described. The recommended query is created by the determination unit 21, the search type determination unit 22, the recommended query creation unit 23, and the recommended query selection unit 24 based on the search query history input by the user.

  The determination unit 21 inputs the first and second search queries arranged in time series, and compares the constituent words of the first and second search queries to determine the change type. In the first embodiment, four types of “restriction”, “generalization”, “relation”, and “switching” are used as change types. Specifically, when there is a change in which the second search query is added to the first search query and the words constituting the search query are “narrowed”, the words constituting the search query are reduced. “Generalization” if there is a change, if some of the words that make up the search query have changed, or if the words that make up the search query have changed and the words are related If neither “relevant” nor the above applies, it is determined as “switching”.

  For example, when the first and second search queries are “Tokyo Station” and “Tokyo Station Restaurant”, the second search query is obtained by adding the word “restaurant” to the first search query “Tokyo Station”. Therefore, it is determined as “narrow down”. When the first and second search queries are “Tokyo Station Restaurant” and “Tokyo Station”, the second search query deletes the word “restaurant” from the first search query “Tokyo Station Restaurant”. Therefore, it is determined as “generalization”. If the first and second search queries are “Tokyo Station Restaurant” and “Tokyo Station French”, respectively, it is determined as “related” because a part of the search query is changed. When the first and second search queries are “Tokyo Station” and “Otemachi”, respectively, the words of the search query are changed, but they are related to each other, so it is determined as “related”. When the search query word is changed, the relevance between the two search queries is calculated using the existing word relevance technology, and it is determined as “relevant” or “switched” based on the calculated relevance. To do.

  The search type determination unit 22 inputs the latest two search queries from the user, and acquires the change type of the user's search behavior using the determination unit 21. The acquired change type is transmitted to the recommended query selection unit 14.

  The recommended query creation unit 23 creates a plurality of recommended query candidates corresponding to a newer search query (referred to as “last search query”) among the latest two search queries input from the user. Then, for each of the created recommended query candidates, the last search query and each recommended query candidate are input to the determination unit 21, and the change type when the search query is changed from the last search query to the recommended query candidate is set. get. The change type is assigned to each recommended query candidate and is transmitted to the recommended query selection unit 24 as recommended query candidate information. FIG. 6 shows an example of recommended query candidate information. Note that existing query recommendation techniques (for example, Patent Document 1 and Non-Patent Document 1) can be used for creating recommended query candidates. At this time, a wide variety of recommended queries may be collected using a plurality of query recommendation techniques.

  The recommended query selection unit 24 holds a transition table, refers to the transition table, and the distribution ratio of the change type given to the recommended query candidate to be selected indicates the change type received from the search type determination unit 22 as the preceding change type. A predetermined number of recommended queries are selected so as to be equal to the distribution ratio of the subsequent change type. The recommended query selection unit 24 creates a transition table from the aggregation table received from the query change aggregation unit 25. The transition table is created by taking the sum of the values of the subsequent change type and dividing the value of the subsequent change type by the sum for each of the preceding change types in the aggregation table. FIG. 7 shows an example of the transition table. In the transition table shown in the figure, four change types are used, and the ratio of the subsequent change types that have transitioned following the preceding change type is described. For example, when the change type of the search behavior of the user is “refining”, the recommended query to which each change type of “switching”, “relevant”, “generalization”, “refining” is assigned is 0.087. , 0.609, 0.217, 0.065, the recommended query is selected so as to approach.

  Next, the flow of processing for creating a recommendation query will be described. FIG. 8 is a flowchart showing a flow of processing for creating a recommendation query.

The search type determination unit 22 inputs the latest two search queries from the user, and acquires the change type of the user's search behavior using the determination unit 21 (step S201). Then, the recommended query creating unit 23 that transmits the acquired change type to the recommended query selecting unit 14 inputs the latest search query, and creates a recommended query candidate corresponding to the search query (step S202).

  Then, a change type in the case of changing from the latest search query to each recommended query candidate is acquired (step S203), and the change type is assigned to each recommended query candidate and transmitted to the recommended query selection unit 24 (step S204). In order to acquire the change type, the determination unit 21 is used similarly to the search type determination unit 22.

  The recommended query selection unit 24 acquires, from the transition table, the ratio of the subsequent change type when the change type received from the search type determination unit 22 is the preceding change type (step S205). For example, when the transition table is as shown in FIG. 7 and the change type received from the search type determination unit 22 is “narrow down”, “switch” is 0.087 and “relevant” is 0.609. , “Generalization” is 0.217, and “Refinement” is 0.065.

  Then, the recommended query selection unit 24 selects and outputs a predetermined number of recommended queries so that the ratio of the change type given to the selected recommended query is equal to the ratio acquired from the transition table (steps S206 and S207). ). For example, the distribution of “switching” is 0.087, “relation” is 0.609, “generalization” is 0.217, and “refining” is 0.065, and the number of recommended queries to be selected is set to 10. In the case of the recommended query candidates, one recommended query to which the “switching” change type is assigned, six recommended queries to which the “related” change type is assigned, and the “generalized” change type. Are selected, and one recommended query to which the change type “squeeze” is assigned is selected and output as a recommended query.

[Example 2]
FIG. 9 is a functional block diagram illustrating the configuration of the search query recommendation device according to the second embodiment.

  The search query recommendation device 3 illustrated in FIG. 9 includes a determination unit 31, a search type determination unit 32, a recommended query creation unit 33, a recommended query selection unit 34, a query change aggregation unit 35, and a learning unit 37. A search query log database 36 is connected to the query change totaling unit 35, and a correct query change database 38 is connected to the learning unit 37. In addition, each part with which the search query recommendation apparatus 3 is provided is good also as what is comprised by the computer provided with the arithmetic processing unit, the memory | storage device, etc., and the process of each part is performed by a program. This program is stored in a storage device included in the search query recommendation device 3, and can be recorded on a recording medium such as a magnetic disk, an optical disk, or a semiconductor memory, or provided through a network.

  In the second embodiment, the learning unit 37 learns a change type determination model for classifying the change type of the search query, and applies the change type determination model to the determination unit 31 to classify the change type of the search query. Various classifications other than “narrow down”, “generalization”, “relation”, and “switching” shown in the first embodiment are used.

  First, the processing of the learning unit 37 will be described.

  The correct query change database 38 stores a plurality of correct query change information in which a change type is assigned to a set of a preceding query and a subsequent query. FIG. 10 shows an example of correct query change information.

  The learning unit 37 reads the correct query change information from the correct query change database 38, and uses the existing machine learning algorithm such as Support Vector Machine or Decision Tree to determine a change type determination model for classifying the search query change type. learn. The determination unit 31 configures a determination unit that reads the change type determination model and determines the change types of the two input search queries.

  Here, a processing flow of the learning unit 37 will be described. FIG. 11 is a flowchart showing a flow of processing in which the learning unit 37 learns the change type determination model.

  First, the learning unit 37 reads one correct query change information from the correct query change database 38, and calculates a vector expression representing the meaning of the phrase for the preceding query and subsequent query of the read correct query change information (step S301).

Subsequently, using the preceding query and the succeeding query included in the read correct query change information, a feature amount effective for determining the change type of the correct query change information is calculated (step S302). Specifically, the number of common words, the number of common characters, the number of added words, the number of added characters, the cosine similarity between vector expressions, the preceding Calculate skew divergence from query to subsequent query, and skew divergence from subsequent query to preceding query. For example, if the preceding query is “Tokyo Station”, the subsequent query is “Tokyo Station Hotel”, and the respective vector expressions are v ₁ and v ₂ , the number of common phrases is 1, and the number of common characters is 3. The number of words and phrases added is 1, the number of added characters is 3, cosine similarity, and skew divergence is as follows.

  Then, it is determined whether or not there is correct query change information for which the feature quantity is not calculated in the correct query change database 38, and if it exists, the process returns to step S301 (step S303).

  When feature amounts are calculated for all correct query change information, the change type determination model is learned by applying the change type of the correct query change information and the set of calculated feature amounts to an existing machine learning algorithm such as Decision Tree. (Step S304).

  The learned change type determination model is stored in the determination unit 31 (step S305).

  The search type determination unit 32, the recommended query creation unit 33, and the query change totalization unit 35 determine the change type using the determination unit 31 to which the change type determination model is applied. Since the recommended query creation process and the transition table creation process are the same as those in the first embodiment, description thereof will be omitted. The determination unit 31 inputs a search query in chronological order, uses the vector expression of the search query, and determines the change type using the change type determination model learned by the learning unit 37.

  As described above, according to the present embodiment, the determination unit 11 that determines the change type of the search action based on the search query arranged in time series is provided, and the search type determination unit 12 displays the history of the search query. When the input is determined to determine the current change type and the recommended query creation unit 13 creates a recommended query candidate for the most recent search query in the search query history, and the most recent search query is changed to each recommended query candidate The change type is determined and given to each recommended query candidate, the recommended query selection unit 14 holds a transition table indicating the change type transition distribution, acquires the distribution ratio of the change type that transitions from the current change type, and recommends Distribution ratio of change type distribution ratio assigned to the selected recommended query when a predetermined number of recommended queries are selected from the query candidates. And by selecting the close so as to recommend a query, it is possible select and present the query to be recommended depending on the intention to change the user's search query, can be expected accuracy of the recommendation query.

1, 2, 3 ... Search query recommendation device 11, 21, 31 ... Determination unit 12, 22, 32 ... Search type determination unit 13, 23, 33 ... Recommended query creation unit 14, 24, 34 ... Recommended query selection unit 25, 35 ... Query change totaling unit 26, 36 ... Search query log database 37 ... Learning unit 38 ... Correct query change database

Claims (9)

Entering a user's search query history and determining a change type that classifies the change intent of the search query derived from the relevance between the search queries;
Creating a plurality of recommended query candidates for the most recent last search query in the search query history;
Determining a change type when a search query is changed from the last search query to each of the plurality of recommended query candidates, and assigning the change type to each of the plurality of recommended query candidates;
Obtaining a distribution ratio of change types that transition from the change type determined from the history of the search query with reference to a transition table that describes the distribution of change type transitions;
Selecting a recommended query from the recommended query candidates so that the distribution ratio of the change type assigned to the recommended query to be selected is close to the distribution ratio acquired from the transition table;
A search query recommendation method characterized by comprising: As the change type, four types of refinement, generalization, association, and switching are used,
In a search query arranged in chronological order, if a word is added, it is narrowed down, if a word is deleted, generalization is performed, if only a part of the word is changed, or if there is a relationship between the search queries The search query recommendation method according to claim 1, wherein a change type is determined as a relation if there is any, and a change type as a switch if the conditions do not apply. The correct answer change type information from the correct answer change type information storing the correct answer change type information including the first search query, the second search query, and the change type when the first search query is changed to the second search query. And learning a determination model for determining a change type of a search query arranged in time series by a machine learning algorithm,
The search query recommendation method according to claim 1, wherein the change type is determined using the determination model.   A step of reading the search queries in time-series order from the search query storage means for storing the search query history, determining a change type between the read search queries, obtaining a change type transition, and creating the transition table; The search query recommendation method according to any one of claims 1 to 3. A determination means for inputting a search query arranged in a time series and determining a change type that classifies the change intention of the search query derived from the relationship between the search queries,
A search type determination unit that inputs a history of a user search query and determines a current change type using the determination unit;
A plurality of recommended query candidates for the latest search query that is newest in the search query history are created, and the search query is changed from the last search query to each of the plurality of recommended query candidates using the determination unit. Determining a change type, and recommending query creating means for assigning the change type to each of the plurality of recommended query candidates,
The change type distribution ratio that transitions from the current change type is obtained with reference to the transition table that describes the distribution of the change type transition, and the change type distribution ratio assigned to the recommended query to be selected is obtained from the transition table. Recommended query selection means for selecting a recommended query from the recommended query candidates so as to be close to the acquired distribution ratio;
A search query recommendation device characterized by comprising: As the change type, four types of refinement, generalization, association, and switching are used,
In the search query arranged in time series, the determination means narrows down when words are added, generalizes when words are deleted, searches only when some words are changed, or searches 6. The search query recommendation device according to claim 5, wherein when there is a relation between queries, the change type is determined as a relation, and when the relation does not apply, the change type is determined as switching. Correct answer change type storage means for storing correct answer change type information consisting of a first search query, a second search query, and a change type when changed from the first search query to the second search query;
Reading the correct answer change type information from the correct answer change type storage means, and learning means for learning a determination model for determining a change type of a search query arranged in time series by a machine learning algorithm,
The search query recommendation device according to claim 5, wherein the determination unit determines a change type using the determination model. Search query storage means for storing a history of search queries;
Transition table creation means for reading search queries from the search query storage means in chronological order, determining a change type between search queries read using the determination means, obtaining a transition of the change type, and creating the transition table; ,
The search query recommendation device according to any one of claims 5 to 7, further comprising:   A search query recommendation program for causing a computer to execute the search query recommendation method according to claim 1.

Images