JP5250009B2 - Suggestion query extraction apparatus and method, and program - Google Patents

Description

  The present invention relates to a suggestion query extraction apparatus and method, and a program.

  In a conventional Web page search, when a query is input by a user, a search result including a plurality of URLs (Uniform Resource Locators) is presented to the user by a search engine on the Web page.

  Furthermore, in recent Web page searches, not only the presentation of search results but also queries related to the input query are suggested as alternative query candidates. A query suggested as a candidate for an alternative query in such a Web page search is called a “suggestion query”.

  In general, as a suggestion query, a query similar to a query and a constituent element (a word form if a word) is presented. For example, when the user erroneously inputs “hodel” where “hotel” should be input as a query, “hotel” is generally presented to the user as a suggestion query. Those that correct such spelling mistakes can also be considered as a kind of suggestion query.

  Furthermore, although the query and the constituent elements are dissimilar, it is convenient for the user if a query similar in meaning to the query, for example, a so-called synonym or synonym if the query is a word can be presented as a suggestion query. For example, in the above example, it is convenient for the user if the synonym of “hotel” such as “inn” or “inn” can also be presented as a suggestion query.

  In order to appropriately extract a query (synonym, synonym, etc.) having a similar meaning to such a query as a suggestion query, the present inventors have developed a technique related to acquisition of a semantic category by a label propagation method using a search click-through log. Has already been proposed (see Non-Patent Document 1).

  Here, the search click-through is a snippet (the title of the web page that hits the query, the URL of the web page that hits the query, and the snippet indicated by the search result returned by the search engine when the user inputs the query) This means that the user clicks (selects) one of the Web pages by looking at a list including a part of the Web page including the query.

  Such a search click-through is considered to represent the user's intention directly. In other words, even if two or more query components (word forms, etc.) are dissimilar, those that reach the same Web page are likely to be queries entered with the same intention. It is done. In particular, it is considered that two or more queries that reach the same Web page are often synonyms. Therefore, by using a search click-through log in which a query and a URL (click destination URL) of a Web page clicked (selected) are stored in association with each other, a query having a similar meaning to a query input by a user (Synonyms, synonyms, etc.) can be appropriately extracted as a suggestion query.

Mamoru Komachi, Shinpei Makimoto, Kei Utsumi, Manabu Sasano, “Semantic Category Acquisition by Label Propagation Using Web Page Search Log”, Research Report Spoken Language Information Processing (SLP), 2009-SLP-76, 9 No. 1-6 pages, May 4, 2009

  However, in the search click-through log, there is a click destination URL that co-occurs with a large number of queries, that is, a so-called generic pattern. For this reason, a phenomenon occurs in which queries having a low semantic similarity are evaluated to have higher semantic similarity than the original through a generic pattern.

  When such a phenomenon occurs, so-called semantic drift occurs, and the accuracy of extracting a suggestion query deteriorates. In this regard, according to Non-Patent Document 1, in the label propagation method, the instance score vector has a parameter α∈ (0, 1) indicating whether the seed label or the graph structure is important, and the parameter α is close to 0. If the parameter α is close to 1, the result is that the graph structure created from unlabeled data is taken into consideration. By adjusting this parameter α, it is possible to suppress the occurrence of semantic drift to some extent. However, when a query co-occurs with only a few click destination URLs including a generic pattern, the occurrence of semantic drift cannot be suppressed even if the parameter α is adjusted.

  Therefore, the present invention provides a suggestion query extraction apparatus and method for improving the precision of suggestion query extraction by suppressing semantic drift caused by the presence of a generic pattern without adjusting the parameter α of the instance score vector. It aims at providing a program.

  Specifically, the present invention provides the following.

(1) For an input query input as a new query from a user terminal based on a click destination URL indicating a click destination of a search result for a query and a click-through log including a plurality of history information associated with the query A suggestion query extraction device that extracts suggestion queries with similar meanings,
Referring to the click-through log, for each of the queries, frequency counting means for counting the number of the click destination URLs associated with each other as a co-occurrence frequency;
Based on the co-occurrence frequencies tabulated by the frequency tabulating unit, an instance pattern matrix generating unit that generates an instance pattern matrix indicating a relationship between the query as an instance and the click-to URL as a pattern;
Based on the instance pattern matrix generated by the instance pattern matrix generation means, a normalized Laplacian matrix calculation means for calculating a normalized Laplacian matrix indicating the association between the query as the instance and the co-occurrence query as a kernel;
A query when the input query is used as a seed according to a label propagation method using the normalized Laplacian matrix computed by the normalized Laplacian matrix computing unit as a kernel in response to receiving the input query from the user terminal A related query extraction unit that calculates a similarity score between meanings of each other and extracts a query having a high similarity score as a related query with priority.
Out of the related queries extracted by the related query extraction means, extracts the suggestion query for the input query according to the ranking based on the similarity score, and sends a suggestion query transmission means to the user terminal;
With
The instance pattern matrix calculation means includes:
For each element of the instance pattern matrix, normalized self-mutual information calculation means for calculating normalized self-mutual information;
Among the normalized self-mutual information amounts calculated for each element by the normalized self-mutual information amount calculation means, a threshold is set to 0, and elements having a negative normalized self-mutual information amount are fixed to a predetermined value. An edge deleting means for deleting an edge connecting an instance and a pattern in the element by replacing with a value ;
A suggestion query extraction device.

  According to such a configuration of the present invention, the normalized Laplacian matrix is created using an instance pattern matrix based on the search click-through log. Since normalized self mutual information is adopted as each element of this instance pattern matrix, the influence of so-called generic patterns is suppressed, and the intensity of label propagation in the label propagation technique is appropriately determined. Therefore, by applying a label propagation method that uses such a normalized Laplacian matrix as a kernel, queries that are inherently low in similarity in meaning are evaluated as having higher similarity than in the original through a generic pattern. The occurrence frequency of the phenomenon can be suppressed. As a result, semantic drift is suppressed, and the accuracy of extracting related queries, that is, the accuracy of extracting suggestion queries can be improved.

(2) a likelihood calculating language model creating means for creating a likelihood calculating language model based on a language resource DB including a plurality of the queries;
Likelihood for calculating the likelihood of the related query extracted by the related query extracting means as a likelihood score indicating the likelihood of query based on the likelihood calculating language model created by the likelihood calculating language model creating means. Degree score calculation means,
Reranking means for reranking the related query extracted by the related query extracting means based on the likelihood score calculated by the likelihood score calculating means in addition to the similarity;
Further comprising
The suggestion query transmission means extracts the suggestion query according to the result of reranking by the reranking means, and transmits it to the user terminal.
The suggestion query extraction device according to (1).

  According to such a configuration of the present invention, the result of reranking based on the likelihood score is used to extract the suggestion query, so that the accuracy of the suggestion query extraction is further improved.

In calculating the likelihood score, the language resource DB and the likelihood calculating language model need only be able to calculate what character or word is likely to be generated as a query based on the distribution of characters and words. Various things can be adopted. Specifically, various types such as a character Ngram language model based on a character-based language resource DB and a word Ngram language model based on a word-based language resource DB can be adopted.
The likelihood can be expressed using a probability distribution such as the appearance frequency of characters or words. However, from the viewpoint of preventing underflow in floating-point arithmetic, natural log likelihood is preferably employed for operation. .

  Furthermore, the present invention provides a method and a program corresponding to the apparatus according to (1). Thereby, the same effect as (1) can be expected.

  According to the present invention, it is possible to improve the accuracy of extracting a suggestion query by suppressing a semantic drift caused by the presence of a generic pattern.

It is a functional block diagram which shows the functional structure of one Embodiment of the information processing system containing the suggestion query extraction apparatus which concerns on this invention. It is a figure explaining the label propagation method employ | adopted as the related query extraction part of the suggestion query extraction apparatus of FIG. It is a figure explaining the label propagation method which uses a normalized Laplacian matrix as a kernel. FIG. 2 is a functional block diagram showing details of a functional configuration of a preparation unit for generating a normalized Laplacian matrix as a kernel in the suggestion query extraction device of FIG. 1. It is a flowchart which illustrates the suggestion query extraction process which the suggestion query extraction apparatus of FIG. 1 performs. 6 is a flowchart illustrating a normalized Laplacian matrix creation process in the suggestion query extraction process of FIG. 5.

  Hereinafter, embodiments of the present invention will be described. This is merely an example, and the technical scope of the present invention is not limited to this.

  This embodiment is applied to a computer and its peripheral devices. Each unit in the present embodiment is configured by hardware and software that controls the hardware provided in the computer and its peripheral devices.

  The hardware includes a storage unit, a communication device, a display device, and an input device in addition to a CPU (Central Processing Unit) as a control unit. Examples of the storage unit include a memory (RAM: Random Access Memory, ROM: Read Only Memory, etc.), a hard disk drive (HDD: Hard Disk Drive), and an optical disk (CD: Compact Disc, DVD: Digital Versatile Drive, etc.). It is done. Examples of the communication device include various wired and wireless interface devices. Examples of the display device include various displays such as a liquid crystal display and a plasma display. Examples of the input device include a keyboard and a pointing device (mouse, tracking ball, etc.).

  The software includes a computer program and data for controlling the hardware. The computer program and data are stored in the storage unit, and are appropriately executed and referenced by the control unit. The computer program and data can be distributed via a communication line, or can be recorded on a computer-readable medium such as a CD-ROM and distributed.

  FIG. 1 is a functional block diagram showing a functional configuration of an embodiment of an information processing system including a suggestion query extraction device according to the present invention.

  The information processing system is configured by connecting a suggestion query extraction device 11 and a user terminal 12 to each other.

  In addition, although the connection form of the suggestion query extraction device 11 and the user terminal 12 is not particularly limited, in the present embodiment, the suggestion query extraction device 11 and the user terminal 12 are connected via the Internet (not shown). To do. Further, although there may actually be a plurality of user terminals 12, here, it is assumed that there is one user terminal 12 for convenience of explanation.

  The suggestion query extraction device 11 includes a main processing unit 21 and preparation units 22 and 23.

  The main processing unit 21 extracts a suggestion query based on a query input from the user terminal 12 (hereinafter referred to as “input query”), and transmits it to the user terminal 12. Therefore, the main processing unit 21 includes a related query extraction unit 31, a likelihood score calculation unit 32, a query list reranking unit 33, and a suggestion query transmission unit 34.

  The related query extraction unit 31 extracts and lists one or more queries related to the input query (hereinafter referred to as “related queries”). Such a list including one or more related queries is hereinafter referred to as a “related query list”.

  As a related query extraction method by the related query extraction unit 31, in this embodiment, according to a label propagation method using a normalized Laplacian matrix as a kernel, the similarity between the queries when the input query is used as a seed is calculated, A technique of extracting a related query based on the similarity is employed. Details of the normalized Laplacian matrix and the label propagation method will be described later.

  In this case, the related query extraction unit 31 can also rank (rank) each of one or more related queries based on the semantic similarity. Here, if a value indicating the level of similarity in meaning is hereinafter referred to as a “similarity score”, each of the one or more related queries is added with a similarity score and sorted in order of ranking. It becomes. In this way, a related query list with a similarity score is generated and held in the related query list holding unit 35.

  The likelihood score calculation unit 32 calculates, for each of one or more related queries included in the related query list, a natural log likelihood as a likelihood score indicating the likelihood of query based on the character Ngram language model. Details of the character Ngram language model will be described later.

  Each likelihood score calculated by the likelihood score calculation unit 32 is associated with each related query and added to the related query list. That is, a related query list with a likelihood score and a similarity score is created and held in the related query list holding unit 35.

  The query list reranking unit 33 calculates the sum of the logarithm of the similarity score and the likelihood score for each of one or more related queries included in the related query list, and based on each calculation result, the one or more related queries Perform query re-ranking (re-ranking). Then, in the related query list with the likelihood score and the similarity score, each of the one or more related queries is re-sorted in the reranking order.

  The suggestion query transmission unit 34 preferentially extracts a high-order related query as a suggestion query from the re-sorted related query list after the reranking, and transmits it to the user terminal 12.

  As described above, the related query list holding unit 35 holds a related query list with a similarity score and a related query list with a likelihood score and a similarity score. The related query list with similarity score and the related query list with likelihood score and similarity score may be held as separate lists, but may be held as one list. Here, holding as one list means that a related item with likelihood score and similarity score is added to the related query list with similarity score by adding an item for storing the likelihood score for each related query. It means to keep as a query list.

  The outline of the functional configuration of the main processing unit 21 of the suggestion query extraction device 11 has been described above. Furthermore, with reference to FIG.2 and FIG.3, the detail of the related query extraction part 31 especially among the main process parts 21 is demonstrated below.

  FIG. 2 is a diagram for explaining a label propagation method employed in the related query extraction unit 31, and is a diagram illustrating a state of label propagation when the seed query relates to travel.

  In FIG. 2, a node indicated by a circle on the left side indicates a query (only a word in the example of FIG. 2). A node indicated by a circle on the right side indicates a pattern that co-occurs with the query on the left side. As described above, the graph shown in FIG. 2 is a bipartite graph in which the left node is a query and the right node is a pattern that co-occurs with the query. In the graph, the strength of the line connecting the left and right nodes (the thick straight line is the strongest in the figure, and the smaller the line, the weaker the shorter the length of the dotted line portion), It shows the degree of co-occurrence between the left and right nodes. The line connecting the left and right nodes is also called “edge”. Further, the darkness of each node (the darkness of the color in the circle in the figure) represents the strength of the relationship with the seed query.

  Here, a URL shown as a pattern (actually, a URL such as “http: // ...”) means a click destination URL. That is, in the present embodiment, not only a query log conventionally used but also a search click-through log is employed as a pattern in order to perform highly accurate learning regarding the calculation of the strength related to the seed query.

  In FIG. 2, it is assumed that the upper left node is a word as a seed query (hereinafter referred to as “seed word”) “airline company A” and is given a predetermined label. In this case, the label attached to the seed word “airline A” is propagated to the pattern “URL: Chubu” which has a high degree of co-occurrence with the seed word “airline A”. Here, the pattern “URL: Chubu departure” indicates that a Web page including the content that the plane departure / arrival place is the Chubu airport in Japan is the click destination URL. Such a pattern “URL: Chubu” has a strong relationship with the seed query, and the label attached to the seed word “airline A” is propagated.

  On the other hand, the pattern “URL: tour” indicates that a Web page introducing a predetermined tour in which singer B is appointed as a commercial performer is a click-to URL. In this case, the pattern “URL: tour” is a relatively neutral pattern because it co-occurs with a query different from the seed query of the word “singer B”.

  Since the word “travel company C” shares the pattern “URL: Chubu departure” and the pattern “URL: tour” with the seed word “airline A”, it was attached to the seed word “airline A”. The label is propagated. The word “travel agency C” to which the label has been propagated in this way is classified as a word that is strongly related to the seed query.

  As described above, the label propagation method refers to a method of sequentially propagating labels attached to nodes given as seeds to adjacent nodes. In the label propagation method, the optimum label is given as a label in a state where the label propagation process has converged.

  In this embodiment, as such a label propagation method, a method using a normalized Laplacian matrix as a kernel is employed. Therefore, a label propagation method using a normalized Laplacian matrix as a kernel will be described below with reference to FIG.

  FIG. 3 is a diagram for explaining a label propagation method using a normalized Laplacian matrix as a kernel.

  As shown in FIG. 3, in the label propagation method using a normalized Laplacian matrix as a kernel, a seed instance vector F (0) and an instance similarity matrix A are given as inputs. Further, an instance score vector F (t) is obtained as an output at the t-th step in learning (t is an integer value of 1 or more).

Here, a set of all instances is represented as χ. An instance means a node on the left side in FIG. 2, that is, a query (word or the like). When learning about the strength of association with a certain seed query, for example, when learning about the strength of association with a trip related to a seed query in the example of FIG. 2, the instance score vector F (t ) Is represented as a vector whose number of dimensions is the number of elements | χ | As an element value of the i-th dimension (i is an integer value in a range of 1 to | χ |) of the instance score vector F (t), how much the instance x _{i of the} set χ is related to the seed query. (In the example of FIG. 2, a score indicating how related to travel is) is adopted. That is, the score indicating the association degree between the seed query instance x _i of the set χ becomes the element values of the i-th dimension instance score vector F (t).

Therefore, when learning about the strength of association with a certain seed query, the seed instance vector F (0) given as an input is a vector having the following element values. That is, in the seed instance vector F (0), when the instance x _i is included in the set of instances given as seeds (input query for the related query extraction unit 31 in FIG. 1), the element value of the i-th dimension Becomes “1”, and element values of other dimensions become “0”.

・・・（１）
インスタンスパターン行列Ｗとは、例えば、インスタンスｘ_ｉとパターンｐ_ｊの関連性を示す値（従来は単純な共起回数であり、本実施形態では後述する正規化自己相互情報量）をｉ行ｊ列の要素値として有する行列をいう。ここで、従来においては、インスタンスパターン行列Ｗは、次の式（２）によって正規化された上で、式（１）に代入されていた。

・・・（２）
ここで、行列Ｄ（Ｎ）は、次の式（３）によって定まる行列Ｎの次数対角行列をいう。

・・・（３） An instance similarity matrix A given as an input is calculated by the following equation (1) using the instance pattern matrix W.

... (1)
The instance pattern matrix W is, for example, a value indicating the relationship between the instance x _i and the pattern p _j (previously a simple number of co-occurrence, normalized self-mutual information amount described later in the present embodiment) i row j A matrix having column element values. Here, conventionally, the instance pattern matrix W is normalized by the following equation (2) and then substituted into the equation (1).

... (2)
Here, the matrix D (N) is an order diagonal matrix of the matrix N determined by the following equation (3).

... (3)

  When learning about the strength of association with a certain seed query, a seed instance vector F (0) and an instance similarity matrix A are given as inputs, and processing according to the procedure of FIG. An instance vector F (t) is output.

・・・（４）
なお、本実施形態では、後述するように、正規化ラプラシアン行列Ｌは、図１の正規化ラプラシアン行列作成部４２によって作成されて、正規化ラプラシアン行列保持部４３に保持される。 That is, as shown in step S1 of the procedure in FIG. 3, a normalized Laplacian matrix L shown in the following equation (4) is created.

... (4)
In the present embodiment, as will be described later, the normalized Laplacian matrix L is created by the normalized Laplacian matrix creation unit 42 in FIG. 1 and held in the normalized Laplacian matrix holding unit 43.

・・・（５） Next, as shown in step S2 of the procedure of FIG. 3, the process of obtaining the instance vector F (t + 1) of the t + 1 step using the calculation result of the t step by the calculation of the equation (5) is incremented by one. It is executed repeatedly every time. Then, the calculation result of Expression (5) at the converged stage is incremented as t = t + 1, and then output as an instance vector F (t).

... (5)

  The instance vector F (t) output in this way is a vector in which instances (queries) are arranged in order of similarity of meaning with respect to the instance given as a seed.

  Therefore, the related query extraction unit 31 (FIG. 1) uses the input query supplied from the user terminal 12 as a seed, executes the above-described steps S1 and S2, and calculates the instance vector F (t). Related queries can be extracted. That is, the related query extraction unit 31 is based on the instance vector F (t), and has the highest first to Kth (S is an integer value of 1 or more) meaning similarity to the input query, that is, 1 to K. Instances corresponding to each element of the dimension can be extracted as K related queries.

  In this case, the 1 to K-dimensional element values of the instance vector F (t) are employed as similarity scores added to each of the K related queries. That is, the repetitive calculation of Expression (5) in step S2 described above is equivalent to performing ranking (ranking) based on the similarity score for each instance (each query) and sorting in order of the ranking results. Therefore, the related query extraction unit 31 can create a related query list with a similarity score by extracting each element of 1 to K dimensions of the instance vector F (t).

  In the equation (5), the parameter α is a parameter indicating which one of the label labeling method and the label propagation method attaches importance to the graph structure, and is variable within a range of 0 to 1. That is, the closer the parameter α is to 0, the more biased the seed label is, and the closer α is to 1, the result is a result of considering a graph structure created from unlabeled data (instances).

Further, when learning about the strength of association with two seed queries, a value of “1” or “−1” is given to each instance given as a seed, so that a seed instance vector F ( 0) is created. Then, the label of the instance x _i is determined by the sign of the final score y _i . Furthermore, when learning about the strength of association with three or more n seed queries, an n-dimensional matrix is created as a seed, not a vector, and labeling is performed.

  Next, a method for creating a normalized Laplacian matrix used as a kernel in such a label propagation method will be described with reference to FIG.

  FIG. 4 is a functional block diagram showing details of the functional configuration of the preparation unit 22 for generating a normalized Laplacian matrix as a kernel in the suggestion query extraction device 11 of FIG.

  The preparation unit 22 includes a click-through log DB 41, a normalized Laplacian matrix creation unit 42, and a normalized Laplacian matrix storage unit 43.

  The click-through log DB 41 stores a search click-through log. That is, the click-through log DB 41 stores a plurality of click destination URLs indicating click destinations of search results for the query and history information associated with the query.

  The normalized Laplacian matrix creation unit 42 includes a co-occurrence frequency counting unit 61, an instance pattern matrix generation unit 62, and a normalized Laplacian matrix calculation unit 63.

The co-occurrence frequency totaling unit 61 refers to the search click-through log from the click-through log DB 41 and totals the number of click destination URLs associated with each query. Here, the number of clicks destination URL that has been aggregated by the co-occurrence frequency totaling unit 61 and queries the instance x _i in χ set above, corresponds to the co-occurrence number w _ij of clicks destination URL as a pattern p _j . Therefore, the number of click destination URLs counted by the co-occurrence frequency counting unit 61 is hereinafter referred to as “co-occurrence frequency”.

  The instance pattern matrix generation unit 62 calculates an instance pattern matrix indicating the association between the instance (query) and the pattern (click destination URL) based on the co-occurrence frequencies counted by the co-occurrence frequency counting unit 61.

  The normalized Laplacian matrix computing unit 63 computes the normalized Laplacian matrix by computing the above-described equation (4) using the instance pattern matrix.

  The normalized Laplacian matrix holding unit 43 holds the normalized Laplacian matrix created by the normalized Laplacian matrix creating unit 42 as a kernel.

Note that the instance similarity matrix A necessary for the normalized Laplacian matrix is calculated according to the equation (1) as described above. However, since it is a very large matrix, the storage capacity may be very large. In such a case, the normalized Laplacian matrix holding unit 43 holds only instances pattern matrix W and its transposed matrix W ^T, normalized Laplacian matrix calculator 63, by calculating the equation (1) each time , Storage capacity can be reduced. This is because the instance similarity matrix A is a dense matrix, whereas the instance pattern matrix W is a sparse matrix.

  Further, an instance pattern matrix necessary for creating a normalized Laplacian matrix as a kernel will be described below.

  As described above in the [Background Art] field, there is a generic pattern that co-occurs with a large number of queries in the click destination URL. For this reason, there has conventionally been a phenomenon in which queries having low semantic similarity are evaluated as having higher similarity than the original via a generic pattern.

  In other words, in the label propagation method, a label is propagated from a propagation source instance (query) to a propagation destination instance having a common pattern (click destination URL). In this case, the spread of propagation from the propagation destination instance is considered as the propagation strength. For this reason, the conventional label propagation method has the following first and second features. That is, the first feature is a feature that propagation becomes weak when a propagation destination instance has a large number of patterns. Further, the second feature is a feature that the propagation is strong when the propagation destination instance has only a small amount of pattern. An example in which the second feature appears prominently is a case where the propagation destination instance has only one pattern and is connected to the propagation source instance only by the pattern. In such a case, even if the propagation destination instance has only one generic pattern, it is strongly propagated. Strongly propagated means that even if the propagation source and destination instances are connected by only one generic pattern, that is, the semantic similarity is inherent even if the semantic similarity is low. It means that it will be evaluated as higher.

  Here, the second feature of the conventional label propagation method, that is, the feature of strong propagation when the propagation destination instance has only a small amount of patterns, is caused by the normalization processing of the instance pattern matrix W. .

That is, in the related art, as shown in the above equation (2), the inverse matrix D ⁻¹ (W) of the order diagonal matrix is multiplied by the left side of the instance pattern matrix W, so that the instance pattern matrix W becomes It was normalized. Specifically, each row of the instance pattern matrix W corresponds to each instance (each query), and each element value of a predetermined row indicates the number of times of co-occurrence between the corresponding instance and each pattern (click destination URL) ( The number of clicks). In each row corresponding to each instance, the sum of the element values is normalized so as to be “1”.

  For this reason, conventionally, each element value is small for a row corresponding to an instance co-occurring with many patterns. In addition, for a row corresponding to an instance in which the distribution of co-occurring patterns is biased, the element value corresponding to the pattern that co-occurs is large.

  On the other hand, conventionally, each element value is large for a row in which co-occurring patterns correspond to a small number of instances. As an extreme example, when there is only one co-occurring pattern, the element value corresponding to the pattern is always “1”. The fact that the element value is always “1” does not change even if the pattern is a generic pattern.

  As described above, the conventional instance pattern matrix W normalized by the expression (2) is a row corresponding to an instance having almost no co-occurring pattern other than the generic pattern, and an element value corresponding to the generic pattern. Has rows that are close to "1". Conventionally, a Laplacian matrix L is created from the instance pattern matrix W normalized by the equation (2), and learning is performed according to a label propagation method using the Laplacian matrix L. As a result, an instance (query) that has almost no co-occurring pattern (click-to URL) other than the generic pattern tends to have a high degree of semantic similarity with the instance (seed query) given as a seed. there were. That is, an instance that has almost no co-occurrence pattern other than the generic pattern and an instance given as a seed correspond to queries that originally have low similarity in meaning. There is a tendency that such queries that are originally low in similarity in meaning are evaluated as having higher similarity in meaning through the generic pattern.

  Therefore, in order to suppress the occurrence of such a phenomenon, as shown in FIG. 4, the instance pattern matrix generation unit 62 of the present embodiment includes a normalized self-mutual information calculation unit 71 and an edge cut unit 72. I have.

  The normalized self-mutual information amount calculation unit 71 calculates a normalized self-mutual information amount (NPMI: Normalized Pointe Mutual Information) as each element value of the instance pattern matrix W. Hereinafter, this normalized self-mutual information amount will be described.

・・・（６）
式（６）において、ｉ（ｘ，ｐ）が、インスタンスｘとパターンｐとの自己相互情報量を示している。即ち、式（６）の右辺において、インスタンスｘとパターンｐとが互いに独立であると仮定して求めた確率分布がｐ（ｘ）ｐ（ｐ）であり、実際に観測された確率分布がｐ（ｘ，ｐ）である。式（６）の右辺に示すように、これらの２つの確率分布の情報量の差が自己相互情報量ｉ（ｘ，ｐ）として求められる。 The self mutual information (PMI: Pointwise Mutual Information) before normalization is expressed by the following equation (6).

... (6)
In Expression (6), i (x, p) represents the self-mutual information amount between the instance x and the pattern p. That is, on the right side of equation (6), the probability distribution obtained on the assumption that the instance x and the pattern p are independent from each other is p (x) p (p), and the actually observed probability distribution is p (X, p). As shown on the right side of Equation (6), the difference between the information amounts of these two probability distributions is obtained as the self-mutual information amount i (x, p).

・・・（７） Here, the range that can be taken as the value of the self-mutual information amount i (x, p) is [−∞ to + ∞], and when the two probability distributions match, the self-mutual information amount i (x, p). Becomes 0. Accordingly, when the self mutual information i (x, p) is directly adopted as each element value of the instance pattern matrix W, all the element values that are “0” when the conventional co-occurrence number is used as the element value are all It becomes “−∞”, and the calculation becomes impossible. Therefore, in the present embodiment, as shown in the following equation (7), the self-mutual information amount i (x, p) is normalized, and the resulting normalized self-mutual information amount in (x, p) is obtained. In principle, it is adopted as each element value of the instance pattern matrix W.

... (7)

  As shown in the equation (7), the normalized self-mutual information in (x, p) is normalized by dividing the self-mutual information i (x, p) by (−lnp (x, p)). The range that the value can take is [−1 to +1]. When the probability distribution p (x, p) is 0, the normalized self-mutual information amount in (x, p) is -1. Further, when the probability distributions p (x) and p (p) are independent from each other, the normalized self-mutual information amount in (x, p) is zero. When the instance x and the pattern p co-occur with each other, the normalized self mutual information amount in (x, p) is 1.

  In the present embodiment, the normalized self-mutual information amount calculating unit 71 of the instance pattern matrix generating unit 62 in FIG. 4 performs the normalized self-mutual information amount in () for each element of the instance pattern matrix W according to Expression (7). x, p) is calculated.

・・・（８）
式（８）において、右辺の［α］^ｔｈは、閾値ｔｈ以下の場合、入力値αを削除し（入力値αを入力としてはみずに、出力せず）、閾値ｔｈを超えている場合、入力値αをそのまま出力する関数を意味している。ここで、閾値ｔｈは、半正定値性を満足させるために０以上の値である必要がある。 However, when the normalized self-mutual information amount in (x, p) of the equation (7) is adopted as each element value of the instance pattern matrix W, the semi-definite property is lost, and therefore the normalized Laplacian matrix is used. The applied label propagation method becomes impossible. Therefore, in the present embodiment, each element value w (x, p) of the instance pattern matrix W is calculated according to the following equation (8).

... (8)
In Expression (8), when [α] ^{th on the} right side is equal to or less than the threshold th, the input value α is deleted (the input value α is not regarded as an input and is not output), and when the threshold th is exceeded, This means a function that outputs the input value α as it is. Here, the threshold th needs to be a value equal to or greater than 0 in order to satisfy the semi-definite property.

  For example, when the threshold th is 0, the right side of Equation (8) means that when the normalized self-mutual information amount in (x, p) is a negative value, the negative value is not seen. ing. That is, the fact that the normalized self-mutual information amount in (x, p) is a negative value means that there is a negative correlation between the instance x and the pattern p, and this combination is unlikely to occur. It means that it is not seen.

  In other words, from the viewpoint of the label propagation method, the fact that the normalized self-mutual information amount in (x, p) is a negative value means that the instance x and the pattern p are not easily edged. That is, in the example of FIG. 2, it means that the strength of the line (edge) connecting the left node indicating the instance x and the right node indicating the pattern p is weak. Here, the significance of using the normalized self-mutual information amount in (x, p) is that the strength for propagating the label is appropriately determined. Therefore, since how to stretch the edge is determined from directly observed data, even if the negative normalized self-mutual information amount in (x, p) is deleted, that is, the edge is deleted, the propagation of the label is performed. There is no particular problem in terms of appropriate strength. In addition, for an element whose normalized self-mutual information amount in (x, p) is 0, it can be determined that the instance x and the pattern p are independent from each other. There is no particular problem in terms of making it appropriate.

  In the present embodiment, the edge cut unit 72 of the instance pattern matrix generation unit 62 in FIG. 4 calculates such a formula (8), so that the value of the normalized self-mutual information amount in (x, p) is a threshold value. Edges in elements below th are deleted. That is, among the elements of the instance pattern matrix W, the value of the normalized self-mutual information in (x, p) is the value of the element whose normalized self-mutual information in (x, p) exceeds the threshold th. It is adopted as an element value as it is. On the other hand, the value of the normalized self-mutual information in (x, p) is not adopted as the element value for the element whose normalized self-mutual information in (x, p) is less than or equal to the threshold th. For example, a predetermined fixed value is adopted.

  As described above, the threshold th serving as a reference for deleting an edge needs to satisfy the semi-definite value, and thus a negative value cannot be adopted, but it is not particularly necessary to adopt 0, and an arbitrary value of 1 or less The positive value of can be adopted.

  As described above, in the present embodiment, the instance pattern matrix generation unit 62 including the normalized self-mutual information calculation unit 71 and the edge cut unit 72 described above generates the instance pattern matrix W according to the equations (7) and (8). The calculated value is supplied to the normalized Laplacian matrix calculation unit 63. In principle, each element of the instance pattern matrix W employs a normalized self-mutual information amount (those that exceed the threshold th), and therefore appropriately determines the label propagation strength in the label propagation method. be able to.

  The normalized Laplacian matrix calculation unit 63 calculates the instance similarity matrix A by calculating the above-described equation (1) using the instance pattern matrix W. Then, the normalized Laplacian matrix calculation unit 63 calculates the normalized Laplacian matrix L by calculating Equation (4) using this instance similarity matrix A, and holds it in the normalized Laplacian matrix holding unit 43 as a kernel. Let

  As described above, by applying the label propagation method using the normalized Laplacian matrix L created by the normalized Laplacian matrix creation unit 42 of the present embodiment as a kernel, The occurrence frequency of the phenomenon that the similarity of meaning is evaluated to be higher than the original through the generic pattern can be suppressed. As a result, semantic drift is suppressed, and the accuracy of extracting related queries, that is, the accuracy of extracting suggestion queries can be improved.

The preparation unit 22 that creates the normalized Laplacian matrix L as a kernel in the suggestion query extraction device 11 of FIG. 1 has been described above.
Next, the preparation unit 23 that creates a likelihood calculation language model in the suggestion query extraction device 11 of FIG. 1 will be described.

  The preparation unit 23 includes a language resource DB 51, a likelihood calculation language model creation unit 52, and a likelihood calculation language model holding unit 53. As the language resource DB 51, the likelihood calculation language model creation unit 52, and the likelihood calculation language model holding unit 53, specifically, any character or word is generated as a query based on the distribution of characters and words. Anything can be used as long as it is easy to calculate, and various things can be adopted. For example, various things such as a character Ngram language model based on a character-based language resource DB and a word Ngram language model based on a word-based language resource DB can be adopted. In the following, the explanation will be continued by taking these examples.

  The language resource DB 51 stores a large number of query logs that have been used as queries, that is, so-called query logs.

  The likelihood calculation language model creation unit 52 creates a likelihood calculation language model based on the query log stored in the language resource DB 51. That is, the likelihood calculation language model creation unit 52 converts a character or word w as a query into a sequence of characters or words w = {x [1], x [2],..., X [n]}. A likelihood calculation language model is created by grasping and calculating the natural log likelihood.

More specifically, for example, the likelihood calculating language model creating unit 52
lnP (w)
= ΣlnP (x [i] | {x [i−N + 1],..., X [i−1]})
= Σ {ln (freq ({x [i−N + 1],..., X [i]})) − ln (freq ({x [i−N + 1],..., X [i−1]}) ))}
The natural log likelihood is calculated according to the following formula.
In this embodiment, the natural log likelihood is calculated. However, this is merely an example, and various things that can express query quality can be used.

  The likelihood calculating language model holding unit 53 holds the character Ngram language model created by the likelihood calculating language model creating unit 52.

The functional configuration of the embodiment of the suggestion query providing system according to the present invention has been described above with reference to FIG.
Next, a flow of a series of processes (hereinafter referred to as “suggestion query extraction process”) executed by the suggestion query extraction device 11 in such a suggestion query provision processing system will be described.

  FIG. 5 is a flowchart illustrating a suggestion query extraction process.

  In step S11, the normalized Laplacian matrix creation unit 42 in FIG. 1 refers to the normalized Laplacian matrix holding unit 43 and determines whether or not a normalized Laplacian matrix has been created.

  When the normalized Laplacian matrix has been created, it is determined as YES in Step S11, and the process proceeds to Step S13. In addition, the process after step S13 is mentioned later.

On the other hand, if the normalized Laplacian matrix has not been created, it is determined as NO in step S11, and the process proceeds to step S12.
In step S12, the normalized Laplacian matrix creation unit 42 creates a normalized Laplacian matrix and causes the normalized Laplacian matrix holding unit 43 to hold it as a kernel. Such processing in step S12 is hereinafter referred to as “normalized Laplacian matrix creation processing”. Details of the normalized Laplacian matrix creation process will be described later with reference to FIG.
When the normalized Laplacian matrix creation process in step S12 is executed, the process proceeds to step S13.

  In step S13, the likelihood calculation language model creation unit 52 refers to the likelihood calculation language model holding unit 53 and determines whether or not a likelihood calculation language model has been created.

  When the likelihood calculation language model has been created, it is determined as YES in Step S13, and the process proceeds to Step S15. The processing after step S15 will be described later.

On the other hand, when the likelihood calculation language model has not been created, it is determined as NO in Step S13, and the process proceeds to Step S14.
In step S <b> 14, the likelihood calculation language model creation unit 52 creates a likelihood calculation language model and causes the likelihood calculation language model holding unit 53 to hold it. Thereby, a process progresses to step S15.

In step S <b> 15, the related query extraction unit 31 determines whether an input query is supplied from the user terminal 12.
When the input query is not supplied from the user terminal 12, it is determined as NO in Step S15, and the process returns to Step S15 again. That is, until the input query is supplied from the user terminal 12, the determination process in step S15 is repeatedly executed, so that the suggestion query extraction process enters a standby state.
Thereafter, when an input query is supplied from the user terminal 12, it is determined as YES in Step S15, and the process proceeds to Step S16.

  In step S16, the related query extraction unit 31 creates a related query list with a similarity score. That is, the related query extraction unit 31 calculates the similarity score of the meanings of the queries when the input query is used as a seed according to the label propagation method using the normalized Laplacian matrix created in the process of step S12 as the kernel. And the related query extraction part 31 gives priority to a query with a high similarity score, extracts as a related query with the said similarity score, sorts these by the ranking order based on a similarity score, and attaches a similarity score Create a related query list.

  In step S17, the likelihood score calculation unit 32 calculates a likelihood score for each of the one or more related queries included in the related query list created in the process of step S16, and adds the likelihood score to the related query list. That is, the likelihood score calculation unit 32 calculates the natural log likelihood as a likelihood score indicating the likelihood of a query based on the character Ngram language model created in the process of step S14. Then, the likelihood score calculation unit 32 creates a related query list with a likelihood score and a similarity score.

  In step S18, the query list reranking unit 33 calculates the sum of the logarithm of the similarity score and the likelihood score for each of one or more related queries included in the related query list, and based on each calculation result, Rerank (rerank) one or more related queries. As a result, in the related query list with the likelihood score and the similarity score, each of the one or more related queries is rearranged in the reranking order.

  In step S <b> 19, the suggestion query transmission unit 34 extracts, as priority queries, some related queries that are ranked higher as a result of reranking from the re-sorted related query list after reranking. To the user terminal 12. This completes the suggestion query extraction process.

  Note that the processing in steps S15 to S19 can be executed as long as the normalized Laplacian matrix and the likelihood calculation language model have been created. Therefore, the start timing of the process of step S15 is sufficient if it is after the end of the processes of steps S11 to S14. That is, it is not particularly necessary to immediately start the process of step S15 after the process of steps S11 to S14, and even if the process of step S15 is started after being separated in time. Good.

  In other words, in the suggestion query extraction apparatus 11 of FIG. 1, each of the main processing unit 21, the preparation unit 22, and the preparation unit 23 can execute processing independently and in parallel with each other. Therefore, for example, the preparation unit 22 may appropriately update the normalized Laplacian matrix held in the normalized Laplacian matrix holding unit 43 independently of the suggestion query extraction process. Similarly, for example, the preparation unit 23 may appropriately update the likelihood calculation language model held in the likelihood calculation language model holding unit 53 independently of the suggestion query extraction process.

  Next, the flow of the normalized Laplacian matrix creation process of step S12 in the suggestion query extraction process of FIG. 5 will be described.

  FIG. 6 is a flowchart illustrating the normalized Laplacian matrix creation process.

  In step S31, the co-occurrence frequency totaling unit 61 of the normalized Laplacian matrix creation unit 42 in FIG. 4 totals the co-occurrence frequencies based on the search click-through log. That is, the co-occurrence frequency totaling unit 61 refers to the search click-through log from the click-through log DB 41 and totals the number of click destination URLs (search click throws) associated with each query as the co-occurrence frequency. .

  In step S32, the instance pattern matrix generation unit 62 generates an instance pattern matrix W based on the co-occurrence frequencies tabulated in the process of step S31.

  Specifically, the normalized self-mutual information amount calculation unit 71 of the instance pattern matrix generation unit 62 performs the normalized self-mutual information amount in (x) for each element of the instance pattern matrix W according to the equation (7) described above. , P). Next, the edge cut unit 72 uses the threshold th (for example, th = for example) among the normalized self-mutual information amount in (x, p) calculated for each element of the instance pattern matrix W according to the equation (8) described above. 0) Delete the following elements. Thereby, the edge of the instance x and the pattern p in the deleted element is deleted. When the instance pattern matrix W is thus calculated, the process proceeds to step S33.

  In step S33, the normalized Laplacian matrix calculation unit 63 calculates the instance similarity matrix A by substituting the instance pattern matrix W calculated in the process of step S32 into the equation (1), and the instance similarity matrix A Is substituted into Equation (4) to calculate the normalized Laplacian matrix L.

  The calculated normalized Laplacian matrix L is held in the normalized Laplacian matrix holding unit 43. Thus, the normalized Laplacian matrix creation process ends. That is, the process of step S12 in FIG. 5 ends, and the process proceeds to step S13.

  In this way, the normalized Laplacian matrix L is created using the instance pattern matrix W based on the search click-through log by the normalized Laplacian matrix creation process. For each element of the instance pattern matrix W, a normalized self-mutual information amount is adopted in principle, so that the intensity of label propagation in the label propagation method is appropriately determined.

  Therefore, by applying a label propagation method using such a normalized Laplacian matrix L as a kernel, it is evaluated that the queries having originally low semantic similarity are higher than the original through the generic pattern. The occurrence frequency of such a phenomenon can be suppressed. As a result, semantic drift is suppressed, and the accuracy of extracting related queries, that is, the accuracy of extracting suggestion queries can be improved.

  As described above, in the suggestion query extraction device 11 of FIG. 1, each of the main processing unit 21, the preparation unit 22, and the preparation unit 23 can execute processing independently and in parallel with each other. . Therefore, the normalized Laplacian matrix creation process of FIG. 5 can be executed not only as the process of step S12 in the suggestion query extraction process but also as a process independent of the suggestion query extraction process. For example, even when the normalized Laplacian matrix L held in the normalized Laplacian matrix holding unit 43 is updated, the normalized Laplacian matrix creation process can be executed.

  As mentioned above, although demonstrated using embodiment of this invention, the technical scope of this invention is not limited to the range as described in the said embodiment. Various modifications or improvements can be added to the above embodiment. Embodiments to which such changes or improvements are added are also included in the technical scope of the present invention.

  In the present specification, the step of describing the program recorded on the recording medium is not limited to the processing performed in time series along the order, but is not necessarily performed in time series, either in parallel or individually. The process to be executed is also included.

  Further, in the present specification, the system represents the entire apparatus including a plurality of apparatuses and processing units.

DESCRIPTION OF SYMBOLS 11 Suggestion query extraction device 12 User terminal 21 Main processing part 22 Preparation part 23 Preparation part 31 Related query extraction part 32 Likelihood score calculation part 33 Query list reranking part 34 Suggestion query transmission part 41 Click through log DB
42 Normalized Laplacian Matrix Generation Unit 43 Normalized Laplacian Matrix Holding Unit 51 Language Resource DB
52 Likelihood calculation language model creation unit 53 Likelihood calculation language model holding unit 61 Co-occurrence frequency counting unit 62 Instance pattern matrix generation unit 63 Normalized Laplacian matrix calculation unit 71 Normalized self-mutual information calculation unit 72 Edge cut unit

Claims (4)

Meaning for an input query input as a new query from the user terminal based on a click destination URL indicating a click destination of a search result for the query and a click through log including a plurality of history information associated with the query. A suggestion query extraction device that extracts similar suggestion queries of
Referring to the click-through log, for each of the queries, frequency counting means for counting the number of the click destination URLs associated with each other as a co-occurrence frequency;
Based on the co-occurrence frequencies tabulated by the frequency tabulating unit, an instance pattern matrix generating unit that generates an instance pattern matrix indicating a relationship between the query as an instance and the click-to URL as a pattern;
Based on the instance pattern matrix generated by the instance pattern matrix generation means, a normalized Laplacian matrix calculation means for calculating a normalized Laplacian matrix indicating the association between the query as the instance and the co-occurrence query as a kernel;
A query when the input query is used as a seed according to a label propagation method using the normalized Laplacian matrix computed by the normalized Laplacian matrix computing unit as a kernel in response to receiving the input query from the user terminal A related query extraction unit that calculates a similarity score between meanings of each other and extracts a query having a high similarity score as a related query with priority.
Out of the related queries extracted by the related query extraction means, extracts the suggestion query for the input query according to the ranking based on the similarity score, and sends a suggestion query transmission means to the user terminal;
With
The instance pattern matrix calculation means includes:
For each element of the instance pattern matrix, normalized self-mutual information calculation means for calculating normalized self-mutual information;
Among the normalized self-mutual information amounts calculated for each element by the normalized self-mutual information amount calculation means, a threshold is set to 0, and elements having a negative normalized self-mutual information amount are fixed to a predetermined value. An edge deleting means for deleting an edge connecting an instance and a pattern in the element by replacing with a value ;
A suggestion query extraction device. A likelihood calculating language model creating means for creating a likelihood calculating language model based on a language resource DB including a plurality of the queries;
Likelihood for calculating the likelihood of the related query extracted by the related query extracting means as a likelihood score indicating the likelihood of query based on the likelihood calculating language model created by the likelihood calculating language model creating means. Degree score calculation means,
Reranking means for reranking the related query extracted by the related query extracting means based on the likelihood score calculated by the likelihood score calculating means in addition to the similarity;
Further comprising
The suggestion query transmission means extracts the suggestion query according to the result of reranking by the reranking means, and transmits it to the user terminal.
The suggestion query extraction device according to claim 1. Meaning for an input query input as a new query from the user terminal based on a click destination URL indicating a click destination of a search result for the query and a click through log including a plurality of history information associated with the query. A suggestion query extraction method executed by a suggestion query extraction device that extracts similar suggestion queries of
Referring to the click-through log, for each of the queries, a frequency counting step of counting the number of the associated click destination URLs as a co-occurrence frequency;
An instance pattern matrix generation step for generating an instance pattern matrix indicating a relationship between the query as an instance and the click-to URL as a pattern, based on the co-occurrence frequencies tabulated by the processing of the frequency tabulation step;
Based on the instance pattern matrix generated by the process of the instance pattern matrix generation step, a normalized Laplacian matrix calculation step for calculating a normalized Laplacian matrix indicating the relationship between the query as the instance and the co-occurrence query as a kernel; ,
When the input query is seeded according to a label propagation method using the normalized Laplacian matrix calculated by the processing of the normalized Laplacian matrix as a kernel in response to receiving the input query from the user terminal A related query extraction step of calculating a similarity score of meanings between the queries and preferentially extracting a query having a high similarity score as a related query;
A suggestion query transmission step of extracting the suggestion query for the input query from the related queries extracted by the processing of the related query extraction step according to the ranking based on the similarity score, and transmitting the extraction query to the user terminal; ,
Including
The instance pattern matrix calculation step includes:
For each element of the instance pattern matrix, a normalized self-mutual information amount calculating step for calculating a normalized self-mutual information amount;
Among the normalized self-mutual information amounts calculated for each element by the processing of the normalized self-mutual information amount processing step, a threshold is set to 0, and elements having negative normalized self-mutual information amount are predetermined. An edge deletion step of deleting an edge connecting the instance and the pattern in the element by replacing with a fixed value of
Suggestion query extraction method including Meaning for an input query input as a new query from the user terminal based on a click destination URL indicating a click destination of a search result for the query and a click through log including a plurality of history information associated with the query. A computer that controls a suggestion query extraction device that extracts similar suggestion queries of
Referring to the click-through log, for each of the queries, a frequency counting step of counting the number of the associated click destination URLs as a co-occurrence frequency;
An instance pattern matrix generation step for generating an instance pattern matrix indicating a relationship between the query as an instance and the click-to URL as a pattern, based on the co-occurrence frequencies tabulated by the processing of the frequency tabulation step;
Based on the instance pattern matrix generated by the process of the instance pattern matrix generation step, a normalized Laplacian matrix calculation step for calculating a normalized Laplacian matrix indicating the relationship between the query as the instance and the co-occurrence query as a kernel; ,
When the input query is seeded according to a label propagation method using the normalized Laplacian matrix calculated by the processing of the normalized Laplacian matrix as a kernel in response to receiving the input query from the user terminal A related query extraction step of calculating a similarity score of meanings between the queries and preferentially extracting a query having a high similarity score as a related query;
A suggestion for executing control to extract the suggestion query for the input query from the related queries extracted by the processing of the related query extraction step according to the ranking based on the similarity score, and to transmit the suggestion query to the user terminal A query transmission control step;
Including
The instance pattern matrix calculation step includes:
For each element of the instance pattern matrix, a normalized self-mutual information amount calculating step for calculating a normalized self-mutual information amount;
Among the normalized self-mutual information amounts calculated for each element by the processing of the normalized self-mutual information amount processing step, a threshold is set to 0, and elements having negative normalized self-mutual information amount are predetermined. An edge deletion step of deleting an edge connecting the instance and the pattern in the element by replacing with a fixed value of
A program that executes control processing including

Images