JP6144968B2 - Information presenting apparatus, method, and program - Google Patents

Description

  In recent years, as the arrival of the big data era is said, the amount of data is increasing explosively. On the Internet, in addition to information transmission from companies and people familiar with computers that have been popular in the past, many other media such as SNS (social network service) that allow easy and easy information transmission have been put into practical use. Even those who do not have specialized knowledge began to send information on a daily basis. For this reason, the amount of electronic information generated on the Internet is enormous.

  Also, in the real world, as various sensors become cheaper, smaller, and more sophisticated, mobile terminals such as mobile phones and smartphones that are privately owned by many people are equipped with GPS as standard. The position information of the mobile terminal can be acquired according to the application. In addition, data such as temperature and humidity is constantly acquired by a sensor node provided for commercial use in a farm, a server room, etc., and is used for operation. As described above, sensor information acquired by various sensors is sequentially stored, accumulated, used, and the like.

  Thus, in order to utilize the enormous amount of information generated every day, it is essential to search for information, and it is important to present the search results effectively so that the user can easily understand them. For example, search and presentation of information on the Internet will be described as an example. Internet search engine is Google (registered trademark) web search, image search, YouTube (registered trademark) video search, Microsoft (registered trademark) search engine Bing (registered trademark) web search, image search, etc. As described above, there are a wide range of objects to be searched, such as text, images, and moving images, but the output of search results is often made in units of several tens of cases in a list format. The list format can be said to be a method of presenting search results in a one-dimensional manner in descending order of the score calculated by the search engine based on the relationship between the query and each object. The user should examine the contents from the search results with higher ranks and cancel the confirmation of the search results when the desired information is obtained.

  However, in the presentation of search results in a one-dimensional list format, when the search key such as a keyword included in the query corresponds to not only information related to a desired target but also information related to other targets, There is a problem that a search result for a desired target and a search result for another target are mixed. For example, it is assumed that “a set of html texts on the Internet including a description relating to the person A” is searched from the web using the name of the person A as a keyword, and the list is to be obtained in descending order of score. At this time, if there is another person B who has the same name as the person A, and the “collection of html texts on the Internet including the description about the person B” has a high score, the search result list includes: "The html text on the Internet including the description regarding the person A" and "the html text on the Internet including the description regarding the person B" are mixed.

  In this case, in order to eliminate the information of another person B who has been mixed from the search results, it is necessary to add information that has A but not B as a search key to the query. It becomes necessary to analyze the characteristics of If the user does not have prior knowledge about A and B, the text shown in the list will be read and compared for characteristic analysis, and a lot of time will be consumed. In the worst case, an additional search key that is not appropriate may be selected, and information about the desired A may be excluded from the search result list.

  In order to solve the problem of presenting search results as a one-dimensional list, a method of presenting two-dimensional information has been proposed. Two-dimensional information presentation is a method of outputting a search result in a two-dimensional graph format, for example. For example, when a handwritten numeral image is given as an input (query), among the images in mnistDB (handwritten character recognition database), the neighborhood relationship between an image group determined to have a high similarity to the query and the input image Has been proposed (see Non-Patent Document 1, for example).

  In the technique of Non-Patent Document 1, images having the same label are arranged close to each other, and images having different labels are arranged far from each other. As a result, a cluster of images having the same label is formed in the two-dimensional graph. Is done. Therefore, since clusters of images with different labels are arranged separately on the two-dimensional graph, a set of desired search results can be obtained even when there are multiple images with different labels that have the same score for the same query. And a set of other search results can be solved.

  However, in the technique described in Non-Patent Document 1, in order to grasp the relationship between the query and the search result, such as where the query corresponds to on the two-dimensional graph and what cluster belongs to, the search is performed by the user. It is necessary to compare the auxiliary information (in the case of the non-patent document 1, in the case of the non-patent document 1) and the graph structure with the graph structure.

  As an information presentation method that supports the user so that the relationship between the query and the search result can be efficiently understood, in addition to the two-dimensional positional relationship of the search result, an absolute unique to the search target DB that does not depend on the query A method to perform a three-dimensional visualization of search results by preliminarily calculating an important index for each object in advance and adding height information as the position coordinates of the node representing the object based on the index. It has been proposed (see, for example, Patent Documents 1 to 3).

  For example, taking a word in a document set on the Internet as an example, the height of a word node is calculated based on the number of documents in which the word indicated by the word node appears, and a word indicating a major word appearing in many documents. The altitude of the node (abstract superordinate concept such as “cooking”) is high, and a word-related altitude map in which the node name is described in a large character font is created. In an elevation map like this, if you want to get an overview of a word, look at the node name of the word node located near the top of the mountain to which the word node belongs. You can imagine what it is like. For example, if a user who does not know the word “recipe” sees the word relationship represented by the elevation map as described above, the word node indicating “recipe” has the word node indicating “dish” as a vertex. Since it is located at the foot of the mountain range, it can be understood that “recipe” is a word related to “cooking”.

  In the example of the visualization of the search result by the above-described two-dimensional graph, it is necessary for the user to visually recognize each cluster in comparison with the handwritten image information written together, but by performing the three-dimensional visualization, By using the information held by the node at a high altitude as a clue, it becomes easy to understand the outline of the information of surrounding nodes, that is, the cluster to which the node belongs.

JP 2009-86858 A JP 2009-86859 A JP 2009-28889A

Aoyama Kazuo et al. “Similarity Search and Visualization Using Network Index Structure”, FIT2008 (7th Information Science and Technology Forum)

  However, in the information presentation method based on the three-dimensional visualization of the search results by the conventional method described above, the elevation map is created by determining the height of the node indicating each object regardless of the query before the input of the query. In addition, since it is not the optimum information presentation for each query, there is a problem that sufficient results cannot be expected as support for the user.

  The present invention has been made in view of the above problems, and an object of the present invention is to provide an information presentation apparatus, method, and program capable of performing optimum information presentation for each query.

In order to achieve the above object, an information presentation device according to a first aspect of the present invention provides an object whose similarity with an input query is greater than or equal to a first predetermined value from a plurality of object information including at least one type of feature related to an object. A search unit for searching for a plurality of information, and a two-dimensional plane coordinate of each of the nodes indicating each of the plurality of object information searched by the search unit, the distance between the nodes is between the object information indicated by each node While calculating the distance according to only the relationship, the nodes indicating the object information having high similarity between the features included in the object information are close to each other, and the nodes indicating the object information having a low similarity are arranged far from each other. The height information of each of the nodes is a height corresponding to only the relationship between the object information indicated by the node and the query. And calculating each of the nodes plotted at positions indicated by the two-dimensional plane coordinates and the three-dimensional coordinates indicated by the height information of each of the nodes calculated by the calculation unit, see containing and a presentation unit for presenting a search result at least one characteristic was applied to each of the nodes included in the object information indicated by the calculation unit, the feature between each node included in the object information indicating A partial network structure in which nodes with similarities greater than or equal to the second predetermined value are connected by edges is represented by a spring model, nodes connected by edges are placed close to each other, and nodes not connected by edges are placed far away. Then, the two-dimensional plane coordinates of each of the nodes are calculated so that all the nodes included in the partial network structure have a stable arrangement.

According to the information presentation device of the first invention, the search unit obtains object information whose similarity with the input query is greater than or equal to the first predetermined value from a plurality of pieces of object information including at least one type of feature related to the object. Search multiple times. Next, the calculation unit determines the two-dimensional plane coordinates of each of the nodes indicating each of the plurality of object information searched by the search unit, the distance between the nodes depends only on the relationship between the object information indicated by each node It is calculated that the nodes indicating the object information having high similarity between the features included in the object information are close to each other, and the nodes indicating the object information having a low similarity are arranged far from each other . At this time, the calculation unit represents a partial network structure in which nodes having similarities between features included in the object information indicated by each node are connected to each other by a second edge with a spring model, and the nodes connected by the edge. Place the nodes close to each other, place the nodes that are not connected by edges far away, and calculate the two-dimensional plane coordinates of each node so that all nodes in the partial network structure are stable. . In addition, the calculation unit calculates the height information of each node so that the height information corresponds to only the relationship between the object information indicated by the node and the query. Then, the presentation unit plots each of the nodes at the position indicated by the two-dimensional plane coordinates and the three-dimensional coordinates indicated by the height information of each of the nodes calculated by the calculation unit, and is included in the object information indicated by each of the nodes A search result in which at least one of the above-described features is assigned to each node is presented.

  Thus, in order to express the search result of the object with respect to the query using the two-dimensional plane coordinates according to the relationship between the object information and the three-dimensional coordinates indicated by the height information according to the relationship between the query and the object information. Optimal information can be presented for each query.

  Further, the calculation unit can calculate height information that increases as the similarity between the feature included in the object information and the feature of the query increases. Thereby, the cluster of the object can be grasped by using the height information indicating the relationship between the query and the object as a clue.

In addition, the calculation unit may be arranged so that nodes indicating object information having high similarity between features included in the object information are close to each other and nodes indicating object information having low similarity are arranged far from each other. more plane coordinates in the calculation child can grasp the cluster object more clearly.

  Moreover, the information presentation apparatus according to the first aspect of the invention can be configured to include a clustering unit that separates each of the nodes into a plurality of clusters according to the relationship between the object information indicated by each of the nodes. As a result, it is possible to automatically perform the separation of the cluster of the objects that have grasped the positional relationship of the nodes and the characteristics of the objects given to the nodes.

  Further, the clustering unit can give each of the plurality of clusters a cluster name based on object information indicated by each of the nodes belonging to each of the plurality of clusters. Thereby, the contents of the cluster can be easily grasped.

  In the information presentation device according to the second invention, in the search result indicated by the three-dimensional coordinates, both the search result set for the first query and the search result set for the second query different from the first query. The search results included in the search results are displayed differently for the search results for the first query and the search results for the second query.

  Thereby, even if it is the same object, since it represents as a node from which height differs in each of the search results with respect to a different query, optimal information presentation can be performed for every query.

Further, in the information presentation method according to the third invention, the search unit obtains object information whose similarity with the input query is greater than or equal to a first predetermined value from a plurality of pieces of object information including at least one type of feature related to the object. A plurality of search steps, and a calculation unit that calculates the two-dimensional plane coordinates of each of the nodes indicating each of the plurality of object information searched by the search unit, the distance between the nodes is the distance between the object information indicated by each node The distance is determined according to the relationship only, and the nodes indicating the object information having high similarity between the features included in the object information are close to each other, and the nodes indicating the object information having a low similarity are arranged to be distant from each other. In addition, the height information of each of the nodes is set to a height corresponding to only the relationship between the object information indicated by the node and the query. A step of calculating so that the presentation unit plots each of the nodes at a position indicated by the two-dimensional plane coordinates and the three-dimensional coordinates indicated by the height information of each of the nodes calculated by the calculation unit; in at least the steps of one presenting the search results given to each of the nodes, only including the step of the calculating unit calculates the two-dimensional plane coordinates of the features included in the object information indicated by each of said nodes, The partial network structure in which the nodes having similarities between the features included in the object information indicated by each node are connected to each other by an edge is represented by a spring model, and the nodes connected by the edge are arranged close to each other. Nodes that are not connected at the edge should be placed far away so that all nodes in the partial network structure will be stable. Is a method of calculating the two-dimensional plane coordinates of each of the nodes.

  In the information presentation method according to the fourth aspect of the present invention, in the search result indicated by the three-dimensional coordinates, both the search result set for the first query and the search result set for the second query different from the first query. The search results included in the search result are displayed differently for the search results for the first query and the search results for the second query.

  Moreover, the information presentation program which concerns on 5th invention is a program for functioning a computer as each part which comprises said information presentation apparatus.

  As described above, according to the information presentation apparatus, method, and program of the present invention, the search result of the object for the query, the two-dimensional plane coordinates according to the relationship between the object information, and the relationship between the query and the object information Since the information is expressed using the three-dimensional coordinates indicated by the height information corresponding to the information, an effect that the optimum information presentation can be performed for each query is obtained.

It is a block diagram which shows the structure of the information presentation apparatus which concerns on 1st Embodiment. It is an image figure which shows an example of the object information stored in search object DB. It is an image figure which shows an example of the three-dimensional visualization result of a search result. It is an image figure which shows an example of a control screen. It is a flowchart which shows the content of the information presentation process routine in 1st Embodiment. It is an image figure which shows an example of the three-dimensional visualization result of the search result with respect to keyword "(alpha)." It is an image figure which shows an example of the three-dimensional visualization result of the search result with respect to keyword "(beta)." It is a block diagram which shows the structure of the information presentation apparatus which concerns on 2nd Embodiment. It is the schematic for demonstrating the median centrality regarding the edge of a graph. It is the schematic for demonstrating naming of a cluster. It is a flowchart which shows the content of the information presentation process routine in 2nd Embodiment.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the following, each embodiment will be described with reference to a case where the search target object is “picture book”. In the following embodiments, the term “search” is used as a concept wider than “search”. Here, “search” is a task for finding an object that satisfies all the conditions given as a query in the object group to be searched. On the other hand, “search” is a task for finding an object that satisfies the most conditions when there is no object that satisfies all the conditions given as a query.

<First Embodiment>
The information presentation apparatus 10 according to the first embodiment includes a CPU (Central Processing Unit), a RAM (Random Access Memory), and a ROM (Read Read) that stores an information presentation program for executing an information presentation processing routine described later. Only memory). The CPU functions as the information presentation device 10 by the CPU reading and executing the information presentation program from the ROM which is an internal storage device.

  The information presentation apparatus 10 receives a query for searching for desired information. The information presentation device 10 searches for a plurality of objects corresponding to the input query, and outputs a search result obtained by three-dimensionally visualizing information indicating the searched object.

  As shown in FIG. 1, the information presentation device 10 functionally includes a query feature extraction unit 12, an object search unit 14, a two-dimensional coordinate calculation unit 16, an elevation calculation unit 18, and a three-dimensional visualization result presentation unit 20. It can be expressed in a configuration that includes it. The object search unit 14 is an example of the search unit of the present invention, the two-dimensional coordinate calculation unit 16 and the elevation calculation unit 18 are examples of the calculation unit of the present invention, and the three-dimensional visualization result presentation unit 20 is the present invention. It is an example of a presentation part.

  The query feature extraction unit 12 receives the input query, extracts features from the query, and passes the extracted query features to the object search unit 14.

  For example, the query feature extraction unit 12 can extract the number of appearances of a predetermined word included in text data given as a query in the query as the feature of the query. More specifically, the query feature extraction unit 12 divides text data given as a query into words by a morphological analyzer, and obtains a morphological analysis result in which information such as part of speech is given to each word. The query feature extraction unit 12 refers to information such as the part of speech of the morphological analysis result, and extracts, for example, an independent word as a predetermined word. In addition, you may employ | adopt words other than an independent word as a word to extract. The query feature extraction unit 12 counts the number of appearances of each extracted independent word in the query, and uses the enumeration of each independent word and the number of appearances as a feature of the query.

  For example, when text data “AAA and BBB swims in the sea” (“AAA” and “BBB” are names of characters appearing in a picture book) is input as a query, the query feature extraction unit 12 The independent words “AAA”, “BBB”, “Sea”, and “Swimming” are extracted from the morphological analysis result obtained by morphological analysis of the input query. Further, the query feature extraction unit 12 counts the number of appearances of each extracted independent word in the query, and extracts a feature of “AAA once, BBB once, ocean once, swim once”.

  It should be noted that the query only needs to have a feature that can be compared with the feature included in the object information stored in the search target object DB 30 to be described later, and even if the query exists in the data in the search target object DB 30 It may not be. For example, if the search target object DB 30 stores information on text included in a picture book that is a search target as a feature of the object, the input query only needs to have text. The expression that appears in the picture book stored in the search target object DB 30 may not be used.

  The object search unit 14 acquires object information having a high similarity to the query from the search target object DB 30 using the query features passed from the query feature extraction unit 12, and receives the object information to the two-dimensional coordinate calculation unit 16. hand over.

  Here, the search target object DB 30 stores a plurality of pieces of object information as shown in FIG. Here, the object is an entity indicating a search target, such as a book, music, an image, a video, and a person. The object information is information including at least one type of feature related to the object as described above, and can include various information such as properties and attributes indicated by the object. For example, when the object is an image, information indicated by each item included in metadata such as the image title, size, creation date and time, information indicating the content of the image itself, that is, an image feature amount extracted from the image data Etc. can be included in the object information indicating the image as a feature of the image.

  Here, “having the same type of feature” and “having the same feature” are different concepts. For example, it is assumed that the two objects A and B are an image having a size of 100 × 100 pixels and an image having a size of 200 × 200 pixels, respectively. Both have a plurality of “features of the same type” such as an image title, a size, and an image feature amount. However, when the “size” is focused on, the values are different, so it can be said that the object has different characteristics.

  Each object information is given an identification number (ID) for identifying an object to be searched.

  The example of FIG. 2 is a case where the search target object is “picture book”. For example, the title of the picture book, the cover image, the author, the author of the illustration, the bibliographic information of the publisher, and the independent words appearing in each picture book. Object information including each number of appearances as a feature is stored. Here, as one of the object features other than bibliographic information, the number of appearances for each independent word appearing in the picture book is used. For example, the image feature extracted from the image data indicating the illustration in the picture book The volume, the voice feature amount extracted from the voice data read from the picture book, and the like may be used as the object feature. Also, bibliographic information may include other bibliographic information as a feature.

  The object search unit 14 uses the query features extracted by the query feature extraction unit 12 and the features included in the object information of each object stored in the search target object DB 30 to use the similarity between the query and each object. And a predetermined number (for example, 30) of object information of objects having a high similarity with the query is acquired from the search target object DB 30 in descending order of the similarity.

  The calculation of the similarity between the query and the object will be described by taking a case where the search target object is “picture book” as an example. First, among the features included in the object information of each picture book stored in the search object DB 30, a feature vector of each picture book is created using features that can be compared with the query features extracted by the query feature extraction unit 12. To do. Here, the number of appearances of each independent word appearing in the picture book is used as the feature of the object. Specifically, using the number of appearances for each independent word appearing in each picture book as a feature of the object stored in the search target object DB 30, it appears in all picture books whose object information is stored in the search target object DB 30. For each picture book, a vector having dimensions of the total number of independent words, and elements corresponding to each dimension corresponding to the independent words in the order of ID and the order of appearance from the first page of each picture book. Then, a value tfidf obtained by multiplying the number of appearances tf (term frequency) in each picture book of an independent word corresponding to the dimension of the element by the inverse picture book appearance frequency idf (inverse document frequency) is written in each element of the vector. Finally, the size of the vector is normalized to 1 to obtain a feature vector for each picture book. tfidf can be calculated, for example, as shown in the following equations (1) to (3).

Here, n _{i, j} is the number of appearances of the independent word i in the picture book j. D is the total number of picture books whose object information is stored in the search target object DB 30, and d is the number of picture books in which the independent word i appears among the picture books whose object information is stored in the search target object DB 30.

For example, object information about a total of three picture books is stored in the search target object DB 30 (D = 3), and the object information of each picture book is represented by the following independent words and the number of appearances thereof. Assume that features are included. Note that the picture book whose ID is n is hereinafter referred to as “picture book n”.
Picture Book 1: A 3 times, B 2 times, C 2 times Picture Book 2: A 1 time, C 3 times, D 2 times Picture Book 3: A 2 times, D 1 time, E 2 times

In this case, idf _i (i = A, B, C, D, E) of each independent word is expressed by the following equation (4).

Further, the value of each element of the vector for the picture book 1 is expressed by the following equation (5).

Finally, in order to normalize the vector size to 1, each element is divided by the square root of the sum of squares of each element. The resulting feature vector v ₁ of the picture book 1 is expressed by the following equation (6).

The feature vector v ₂ of the picture book 2 and the feature vector v 3 of the picture book ₃ are created in the same procedure.

Next, a feature vector of the query is created. Assume that the query feature extraction unit 12 extracts, for example, features represented by the following independent words and the number of appearances thereof from the query.
Query: A 2 times, C 3 times

In preparing a feature vector of a query, the value of idf _i in calculating tfidf utilizes values obtained from the search object DB 30, i.e., the expression (4). Others are obtained in the same procedure as the feature vector of the picture book as shown in the following equation (7).

Finally, in order to normalize the magnitude of the vector to 1, each element is divided by the square root of the sum of squares of each element to create a _query feature vector v _query = (0, 0, 1, 0, 0). Independent words that appear in the query and that do not appear in the picture book whose object information is stored in the search target object DB 30 may be ignored. That is, the independent word is not considered as an element of the query feature vector.

Next, the cosine similarity between the feature vector v _query of the created _query and each of the feature vectors v _n (n = 1, 2, 3) of each picture book is calculated. Here, since the size one feature vector is normalized to 1, cosine similarity becomes the inner product equivalent to the v _query and v _n. Also, the inner product with itself becomes 1, which is the maximum value of cosine similarity.

The object search unit 14 acquires, from the search target object DB 30, object information including features whose similarity to the query features is a predetermined value or more. Here, the feature vector v _query and cosine similarity between the feature vector v _n of each object query in descending order, and acquires the object information of the pre-specified predetermined number (e.g., 30 cases). Note that the number of acquired objects is not limited to the upper predetermined number of similarities, and the total number of object information stored in the search target object DB 30 may be specified, and all object information may be acquired in order of similarity.

  Moreover, the object search part 14 calculates the similarity between each object based on at least one of the features included in the acquired object information. The similarity between objects can be calculated by, for example, the cosine similarity of the feature vector of each object calculated above. The object search unit 14 passes the acquired object information, the calculated similarity between objects, and the calculated query and the similarity between each object to the two-dimensional coordinate calculation unit 16.

  In addition, the similarity between each object and the similarity between a query and each object are not limited to the case of obtaining from the similarity of each feature vector. For example, a subjective similarity or dissimilarity can be given in advance for all combinations of two objects whose object information is stored in the search target object DB 30. In this case, given similarity or dissimilarity can be directly used as the similarity between objects. The similarity between the query and each object may be the similarity or dissimilarity given to another object when combined with the object indicated by the object information including the feature that matches the feature of the query.

  In the present embodiment, the object search unit 14 sets a node indicating acquired object information as information indicating the similarity between objects, and the similarity between features included in the object information indicated by each node is a predetermined value. A partial network structure in which the above nodes are connected by an edge is created (see Japanese Patent Laid-Open Nos. 2008-293444 and 2008-305072). The object search unit 14 stores the acquired object information and the similarity between the calculated query and each object in each node constituting the partial network structure. The object search unit 14 transfers the partial network structure thus created to the two-dimensional coordinate calculation unit 16.

  The two-dimensional coordinate calculation unit 16 calculates the two-dimensional plane coordinates of the node indicating each object information based on the object information passed from the object search unit 14 and the similarity between the objects. At this time, two-dimensional plane coordinates are set such that nodes indicating object information having high similarity between features included in the object information are arranged close to each other, and nodes indicating object information having low similarity are arranged distant from each other. calculate.

  In the present embodiment, the partial network structure delivered from the object search unit 14 is used as a spring model with a fixed spring length, and the two-dimensional plane coordinates of a node indicating each piece of object information are calculated. Specifically, in a partial network structure, nodes connected by edges are placed close to each other, nodes not connected by edges are placed far away, and all nodes included in the partial network structure are placed stably. The two-dimensional plane coordinates of each node are obtained so that

  Note that two-dimensional plane coordinates are calculated so that nodes indicating object information with high similarity between features included in the object information are arranged close to each other, and nodes indicating object information with low similarity are arranged far away. Any technique other than the spring model may be used as long as the technique can be used. In addition, as described above, the two-dimensional plane coordinates are not limited to the case where the two-dimensional plane coordinates are calculated using the cosine similarity between the feature vectors using the feature of the object represented by the number of times the independent word appears in the picture book. What is necessary is just to calculate according to the relationship between the object information which a node shows. For example, two-dimensional coordinates are calculated such that nodes indicating objects that match one of the features included in the object information (for example, “author”) are close to each other, and nodes indicating objects having different features are distant from each other. May be.

  The two-dimensional coordinate calculation unit 16 passes the calculated two-dimensional plane coordinates of each node, the object information held by each node, and the similarity between the query and each object to the elevation calculation unit 18.

  The altitude calculation unit 18 calculates the altitude corresponding to the similarity between the query passed from the two-dimensional coordinate calculation unit 16 and each object for the node indicating each object information acquired by the object search unit 14. The altitude is height information for obtaining three-dimensional coordinates in addition to the two-dimensional plane coordinates of the node indicating each piece of object information calculated by the two-dimensional coordinate calculation unit 16.

  As the altitude, the cosine similarity between the feature vector of each object and the query feature vector may be used as it is, or the cosine similarity is multiplied by a predetermined constant (for example, 4) to obtain the height in three-dimensional coordinates. An elevation that emphasizes may be calculated. By emphasizing the height, a cluster to which a desired object corresponding to the query belongs is emphasized. Further, in order to emphasize the cluster to which the desired object belongs, the altitude may be obtained by a function that emphasizes a small value of similarity, such as the logarithm or square root of similarity, or by such a function. The value may be further multiplied by a constant.

  The elevation is not limited to the case where the altitude is calculated using the cosine similarity between the feature vectors using the feature of the object represented by the number of times the independent word appears in the picture book. What is necessary is just to calculate according to the relationship. For example, when a word (for example, an independent word) included in the query is included in any of the object characteristics (for example, “title”), the altitude may be calculated to be higher.

  The altitude calculation unit 18 three-dimensionally calculates the altitude of each node, the two-dimensional plane coordinates of each node passed from the two-dimensional coordinate calculation unit 16, and the similarity between the query held by each node and each object. The result is transferred to the visualization result presentation unit 20.

  The three-dimensional visualization result presentation unit 20 adds the elevation calculated by the elevation calculation unit 18 to the two-dimensional plane coordinate of each node calculated by the two-dimensional coordinate calculation unit 16, and the three-dimensional coordinates of the node indicating each object information , And a three-dimensional visualization result is generated by plotting it on a three-dimensional graph. In addition, the 3D visualization result presentation unit 20 edits the object indicated by each node so that the object indicated by each node can be identified using at least one of the features included in the object information indicated by each node. FIG. 3 shows an example of a three-dimensional visualization result when the search target object is “picture book”. FIG. 3 shows an example in which a cover image of a picture book that is one of the features included in the object information is pasted to each node. Thereby, it is possible to express which picture book each node represents. Moreover, you may use the characteristics other than a cover image, for example, you may use the text which shows the title of a picture book.

  The 3D visualization result presentation unit 20 presents auxiliary information related to the object information indicated by the selected node when a portion of each node is selected on a display device or the like on which the 3D visualization result is displayed. The auxiliary information may include object characteristics, in particular, characteristics indicating bibliographic items, similarity and ranking between the query and each object, contributing words, and the like. A contribution word is a word with a high contribution with respect to having been selected as an object with a high similarity with a query. For example, an independent word corresponding to an element whose product between elements when a cosine similarity between a query feature vector and an object feature vector is calculated can be a contributing word.

  Further, the three-dimensional visualization result presentation unit 20 may highlight a node indicating an object having a high degree of similarity with the query, that is, a node having a higher altitude by adding a darker color to the cover frame. Also, changes in parameters such as the viewpoint of the presented 3D graph, the constants when calculating the altitude, and the spring length and repulsive force of the spring model are accepted by a user's mouse operation or numerical input from the keyboard, etc., for 3D visualization You may change the display of a result and may perform assistance which helps grasp the three-dimensional position of a search result. Thereby, a three-dimensional visualization result can be presented in a state that is easier for the user to see. In addition, the 3D visualization results indicate a grid corresponding to a 2D plane, and the vertex is determined according to the altitude (similarity between the query and each object) of a node having a 2D plane coordinate near the vertex of each grid. The displayed grid may be presented together. As a result, a topographical expression can be added to the three-dimensional visualization result, and the three-dimensional coordinate position of each node can be grasped more intuitively.

  FIG. 4 shows an example of a control screen for changing the display of the auxiliary information and the three-dimensional visualization result. In the example of FIG. 4, on the “information” tab of the control screen, the order of similarity to a query, picture book title, cover image, bibliographic information of an object such as an author, similarity to a query, and contribution words are supported. Information is displayed. Although not shown in detail, the “display setting” tab and the “spring model setting” tab can be used as a screen for changing the various parameters described above. In addition to the above example, auxiliary information and support for helping to grasp the three-dimensional position of the search result may be added.

  In the example of the 3D visualization result in FIG. 3, more specifically, the search target object is “picture book”, and the query “AAA and BBB swims in the sea” (“AAA” and “BBB” are “AAA”. CBB between feature vectors using features represented by the number of appearances of independent words from a picture book whose object information is stored in the search target object DB 30. The search result which acquired 30 picture books with high similarity is shown.

  In FIG. 3, the node group surrounded by the left circle has a higher altitude (similarity to the query) than other nodes, but these are referred to as “AAA” and “BBB” when referring to the auxiliary information. It can be seen that the word groups are picture book groups determined to be similar. That is, it can be said that it is a cluster indicating a picture book group of a series “AAA and BBB” in which characters “AAA” and “BBB” appear. On the other hand, the nodes surrounded by the circle on the right are lower in elevation than other nodes, but these are picture books that are determined to be similar with the words “sea” and “swim” as keys. I know that there is. That is, it can be said that it is a cluster indicating a picture book group including contents related to events at sea. In the example of FIG. 3, a search result having a high degree of similarity with a query can be obtained from the input query with respect to two viewpoints of a picture book of a series and a picture book including contents related to an event at sea.

  The conventional 3D visualization method does not visualize the result according to the query, so it is possible to grasp each of the above-mentioned series of picture book clusters and picture book clusters including contents related to the events in the sea as clusters. However, in this embodiment, since the result is visualized according to the query, the characteristics of the query (in this case, one volume of the series and the ocean It is possible to effectively visualize the stage).

  By presenting such information, it is possible to help grasp a cluster of a plurality of searched objects by using the altitude of the node indicating each object in the three-dimensional visualization result as a clue. Also, by referring to the auxiliary information, it is possible to confirm the meaning of the cluster grasped by the altitude.

  Next, the operation of the information presentation apparatus 10 according to the first embodiment will be described. In the information presentation device 10, an information presentation processing routine shown in FIG. 5 is executed.

  In step 100, the query feature extraction unit 12 receives the input query, extracts a feature from the query, and passes the extracted query feature to the object search unit 14. For example, the query feature extraction unit 12 divides text data given as a query into words by a morphological analyzer, and obtains a morphological analysis result in which information such as part of speech is added to each word. Then, the query feature extraction unit 12 refers to information such as the part of speech of the morphological analysis result, extracts, for example, independent words, counts the number of appearances of each independent word in the query, and each independent word and the number of appearances thereof. The characteristic of the query is the enumeration of

  Next, in step 102, the object search unit 14 acquires object information having a high similarity to the query from the search target object DB 30 using the query features passed from the query feature extraction unit 12. For example, the object search unit 14 uses the feature of the query among the feature vectors created from the query features extracted by the query feature extraction unit 12 and the features included in the object information of each object stored in the search target object DB 30. The cosine similarity with the feature vector created from the features that can be compared with is calculated as the similarity between the query and each object, and the object information having a high similarity between the feature vectors with the query is obtained from the search target object DB 30. A predetermined number (for example, 30) is acquired in descending order of similarity.

  Then, the object search unit 14 calculates the similarity between the objects based on at least one of the features included in the acquired object information, and acquires the acquired object information, the calculated similarity between the objects, and the calculation. The degree of similarity between the query and each object is transferred to the two-dimensional coordinate calculation unit 16. For example, the object search unit 14 sets a node indicating the acquired object information based on the similarity between the acquired object features, and the similarity between the features included in the object information indicated by each node is equal to or greater than a predetermined value. The nodes to be connected are connected by edges, and a partial network structure in which the object information and the similarity between the query and each object are held in each node is created and transferred to the two-dimensional coordinate calculation unit 16.

  Next, in step 104, the two-dimensional coordinate calculation unit 16 calculates the two-dimensional position plane coordinates of the node indicating each object information based on the object information passed from the object search unit 14 and the similarity between the objects. The nodes indicating object information with high similarity between features included in the object information are arranged close to each other, and the nodes indicating object information with low similarity are arranged so as to be distant. For example, the two-dimensional coordinate calculation unit 16 uses the partial network structure delivered from the object search unit 14 as a spring model with a fixed spring length, and nodes connected by edges are arranged close to each other and are not connected by edges. The nodes are arranged far from each other, and the two-dimensional plane coordinates of each node are obtained so that all the nodes included in the partial network structure are stably arranged. Then, the two-dimensional coordinate calculation unit 16 passes the calculated two-dimensional plane coordinates of each node, the object information held by each node, and the similarity between the query and each object to the elevation calculation unit 18.

  Next, in step 106, the altitude calculation unit 18 responds to the degree of similarity between the query passed from the two-dimensional coordinate calculation unit 16 and each object for the node indicating each object information acquired by the object search unit 14. Calculate the altitude. The elevation calculation unit 18 calculates the calculated elevation of each node, the two-dimensional plane coordinates of each node passed from the two-dimensional coordinate calculation unit 16, the object information held by each node, and the similarity between the query and each object. Passed to the three-dimensional visualization result presentation unit 20.

  Next, in step 108, the three-dimensional visualization result presentation unit 20 adds the elevation calculated by the elevation calculation unit 18 to the two-dimensional plane coordinates of each node calculated by the two-dimensional coordinate calculation unit 16, and each object information Is obtained, and a three-dimensional visualization result is generated by plotting it on a three-dimensional graph. Then, the three-dimensional visualization result presentation unit 20 edits the object indicated by each node so that the object indicated by each node can be determined using at least one of the features included in the object information indicated by each node. For example, when the search target object is “picture book”, the cover image of the picture book is pasted to each node. In addition, the 3D visualization result presentation unit 20 adds other auxiliary information and a support function that helps grasp the 3D position of the search result, and outputs the 3D visualization result to be presented on the display device. Then, the information presentation processing routine is terminated.

  As described above, according to the information presentation apparatus according to the first embodiment, the three-dimensional coordinates of the node indicating each of the objects similar to the query, the two-dimensional plane coordinates based on the similarity between the objects, and By expressing the altitude based on the similarity between the query and the object, each object group is separated according to the query compared to the 3D visualization of the conventional method that was uniform for any query. Optimal information can be presented for each query.

  In addition, by including a plurality of features related to the object in the object information, for example, using the features extracted from the content of the object itself, the similarity between the plurality of objects included in the search result, and the similarity between the object and the query And by combining features of objects such as identifying a plurality of objects included in a search result using features indicating bibliographic items of the objects. Can be visualized.

  FIG. 6 and FIG. 7 show an example of the search result for each query when the search target object is “picture book”. The example of FIG. 6 is an example of a search result when the keyword “α” is used as a query. This indicates that the search result is that the altitude of the node indicating the picture book “XXX” is high and the altitude of the node indicating the picture book “YYY” is low. The example of FIG. 7 is an example of a search result when the keyword “β” is used as a query. As in the case of FIG. 6, the search results include the picture books “XXX” and “YYY”, but the elevation of the node indicating the picture book “XXX” is lower than that in FIG. The altitude of the node shown is higher than in the case of FIG.

  As described above, even when the same object is included in search results for different queries, the altitude of the node indicating the object differs depending on the relationship between the query and the object, and the search result corresponds to the query.

<Second Embodiment>
Next, a second embodiment will be described. In the first embodiment, for example, the case where the user grasps the separation of the clusters by observing the three-dimensional visualization result as shown in FIG. 3 and the auxiliary information as shown in FIG. 4 has been described. In this embodiment, a case will be described in which clusters are automatically separated. In the information presentation device according to the second embodiment, the same parts as those of the information presentation device 10 according to the first embodiment are denoted by the same reference numerals, and detailed description thereof is omitted.

  As shown in FIG. 8, the information presentation apparatus 210 according to the second embodiment functionally includes a query feature extraction unit 12, an object search unit 14, a two-dimensional coordinate calculation unit 16, an altitude calculation unit 18, and a clustering. It can represent with the structure containing the part 19 and the three-dimensional visualization result presentation part 20. FIG.

  The clustering unit 19 adds the elevation calculated by the elevation calculation unit 18 to the two-dimensional plane coordinates of each node calculated by the two-dimensional coordinate calculation unit 16, and obtains the three-dimensional coordinates of the node indicating each object information. A node indicating object information is separated into clusters.

  Specifically, the clustering unit 19 introduces the concept of betweenness centerality regarding the edges of the graph. The intermediary centrality regarding the edge of the graph is an idea that, for all combinations of two nodes in the graph, the edge that appears most frequently in the shortest path between the two nodes is the center in the graph. For example, taking the graph shown in FIG. 9 as an example, the shortest path between each two nodes is as follows.

AB: A → B
AC: A → C
AD: A → B → D
AE: A → B → D → E
AF: A → B → D → F
AG: A → B → D → F → G
BC: B → C
BD: B → D
BE: B → D → E
BF: B → D → F
BG: B → D → F → G
CD: C → B → D
CE: C → B → D → E
CF: C → B → D → F
CG: C → B → D → F → G
DE: D → E
DF: D → F
DG: D → F → G
EF: E → F
EG: E → G
FG: F → G

  Of these shortest paths, the most frequently appearing edge is B → D and appears 12 times. By pruning this edge, the graph shown in FIG. 9 can be separated into two clusters of A, B, and C node groups and D, E, F, and G node groups.

  However, even if one edge is removed, it may not be divided into clusters. In this case, after removing the edge, the number of appearances of the edge appearing in the shortest path in each cluster is checked again, and the edge having the maximum mediation centrality is pruned. If it still does not divide into clusters, it continues pruning the edge with the greatest median centrality until it is divided into clusters.

  Actually, it is difficult to determine how many clusters are to be pruned until they are separated, and it can be said that if all the edges are pruned, it is divided into clusters corresponding to the number of nodes. Therefore, for example, a technique called modularity maximization (Reference Non-Patent Document 1 “Newman, M. E. J.“ Modularity and community structure in networks ”. PROCEEDINGS-NATIONAL ACADEMY OF SCIENCES USA 103 (23), 2006”) can be used. In this method, an intuitively “function that increases when there are more edges from one node to a node belonging to the same cluster than an edge to a node belonging to another cluster”, for example ( 8) Modularity Q, which is an index that can be said to be a good cluster separation, is introduced as shown in equation (8), and Q is examined every time an edge is pruned, and Q takes the maximum value. The pruning edge is pruned as many times as the median centrality.

Here, N is the total number of nodes, A is the adjacency matrix of the graph, A _{i, j} is the value of the adjacency matrix between the i-th node and the j-th node, k _i is the number of edges joined to the i-th node, M is the total number of edges in the graph.

  Further, the clustering unit 19 performs automatic annotation, and uses, for example, information that is shared by the object information indicated by the nodes in the cluster as the cluster name. Specifically, the clustering unit 19 acquires features that exist in both the query and the object among the features used when the object search unit 14 calculates the similarity between the query and each object. For example, as described above, when the search target object is a “picture book” and the number of independent words appears as a feature, object information is stored in the query feature vector and the search target object DB 30. Independent words included in common as elements in both feature vectors of the picture book are acquired. Then, in the features included in each of the object information shown by each of all the nodes belonging to each cluster, the number of appearances of the acquired common independent words is accumulated, and the common independent word having the highest number of appearances is The name of

For example, as shown in FIG. 10 and the following, it is assumed that a common independent word exists in both the query and the picture book (object) in the feature included in the object information indicated by each node subjected to cluster separation.
A: α, γ
B: α, λ
C: α, Δ
D: α, β, Δ
E: β, Δ
F: β, η
G: η, ν

  In this case, the cluster of the node group of A, B, and C is named as an independent word “α” having the highest number of appearances. By viewing the cluster name “α” for the clusters of the node groups A, B, and C, the user can easily recognize that the cluster is related to “α”. On the other hand, the cluster of the node group of D, E, F, and G is named as an independent word “β” having the highest number of appearances. Although the self-supporting word “β” does not appear in the picture book indicated by the node G, it belongs to the same cluster as the picture book indicated by the nodes E and F having the self-supporting word “β”. Can be understood as a picture book related to the independent word “β”.

  As described above, the information used for the cluster name is not limited to a common independent word between the query and each object, and information common to the query and other features included in the object information may be acquired. For example, a word that is common to a word included in a query and a word included in the title of a picture book that is one of the features is acquired, and the common word that appears the most in the cluster as described above It may be a cluster name.

  In the above description, the case where the separation of clusters is performed using the technique called modularity maximization has been described. J. Math. 8, 399−404, 1956 ”, spectral clustering (reference non-patent document 3“ U Von Luxburg, “A tutorial on spectral clustering,” Statistics and computing, 2007 ”, topic model“ reference non-patent document 4 “DM Blei, AY Ng, MI Jordan,“ Latent dirichlet allocation, ”the Journal of machine Learning research, 2003” etc. may be used. Although it may be said in various contexts such as community structure detection, the separation of clusters in this embodiment means all of them.

  Next, the operation of the information presentation device 210 according to the second embodiment will be described. In the information presentation device 210, an information presentation processing routine shown in FIG. 11 is executed. In addition, about the process same as the information presentation process routine in 1st Embodiment shown in FIG. 5, the same code | symbol is attached | subjected and detailed description is abbreviate | omitted.

  In steps 100 to 106, an object information group similar to the query is acquired, and the two-dimensional plane coordinates and altitude of the node indicating each object information are calculated.

  Next, in step 200, the clustering unit 19 adds the elevation calculated by the elevation calculation unit 18 to the two-dimensional plane coordinates of each node calculated by the two-dimensional coordinate calculation unit 16, and the tertiary of the node indicating each object. Find original coordinates. Then, for example, by introducing the concept of betweenness centerality regarding the edges of the graph and pruning the edges, each node is separated into clusters.

  Next, in step 202, the clustering unit 19 belongs to the cluster among the information that the object information indicated by the nodes in the cluster has in common, for example, independent words that appear in common in the query and the object. A common independent word having the highest number of appearances in the feature included in the object information indicated by the node is acquired and named as the cluster name of the cluster.

  Next, in step 108, the three-dimensional visualization result is presented together with the cluster separated in step 200 and the cluster name named in step 202, and the information presentation processing routine is terminated.

  As described above, according to the information presentation apparatus according to the second embodiment, it is possible to automatically perform cluster separation of object information and cluster naming, so that the user can easily grasp the separation of clusters. can do.

  The present invention is not limited to the above embodiment, and various modifications and applications can be made without departing from the gist of the present invention.

  In the specification of the present application, the program has been described as an embodiment in which the program is installed in advance. However, the program stored in an external storage device or recording medium is read as needed, and is downloaded and executed via the Internet. You may make it do. In addition, the program can be provided by being stored in a computer-readable recording medium.

DESCRIPTION OF SYMBOLS 10,210 Information presentation apparatus 12 Query feature extraction part 14 Object search part 16 Two-dimensional coordinate calculation part 18 Elevation calculation part 19 Clustering part 20 Three-dimensional visualization result presentation part

Claims (6)

A search unit that searches a plurality of object information having a similarity with the input query from a plurality of object information including at least one type of feature related to the object, the first predetermined value or more;
The two-dimensional plane coordinates of each of the nodes indicating each of the plurality of object information searched by the search unit, the distance between each node is a distance according to only the relationship between the object information indicated by each node, Calculation is performed so that nodes indicating object information with high similarity between features included in the object information are close to each other and nodes indicating object information with low similarity are arranged far apart, and height information of each of the nodes A calculation unit that calculates the height according to only the relationship between the object information indicated by the node and the query;
Each of the nodes is plotted at a position indicated by the three-dimensional coordinates indicated by the two-dimensional plane coordinates and height information of each of the nodes calculated by the calculation unit, and is included in the object information indicated by each of the nodes A presentation unit for presenting a search result obtained by assigning at least one of the nodes to each of the nodes ,
The calculation unit represents a partial network structure in which nodes having similarities between features included in the object information indicated by each node are connected to each other by an edge with a spring model, and the nodes connected by the edge are Nodes that are arranged close to each other and not connected by edges are arranged far from each other, and two-dimensional plane coordinates of each of the nodes are calculated so that all the nodes included in the partial network structure are stably arranged.
Information presentation device.   The information presentation apparatus according to claim 1, wherein the calculation unit calculates height information that increases as a similarity between the feature included in the object information and the feature of the query increases.   The information presentation device according to claim 1, further comprising a clustering unit that separates each of the nodes into a plurality of clusters according to a relationship between the object information indicated by each of the nodes.   The information presenting apparatus according to claim 3, wherein the clustering unit assigns to each of the plurality of clusters a cluster name based on object information indicated by each of the nodes belonging to each of the plurality of clusters. A step of searching for a plurality of object information whose similarity with the input query is a first predetermined value or more from a plurality of object information including at least one type of feature related to the object;
The calculation unit corresponds to the two-dimensional plane coordinates of each of the nodes indicating each of the plurality of object information searched by the search unit, and the distance between the nodes depends only on the relationship between the object information indicated by the nodes. The distance is calculated so that nodes indicating object information having high similarity between features included in the object information are close to each other, and nodes indicating object information having low similarity are arranged far from each other, and each of the nodes Calculating the height information so that the height information depends only on the relationship between the object information indicated by the node and the query;
The presentation unit plots each of the nodes at the position indicated by the two-dimensional plane coordinates and the three-dimensional coordinates indicated by the height information of each of the nodes calculated by the calculation unit, and the object information indicated by each of the nodes Presenting a search result obtained by assigning at least one of the features included in each of the nodes ,
In the step of calculating the two-dimensional plane coordinates by the calculation unit, a partial network structure in which nodes having similarities between features included in the object information indicated by each node are connected to each other by an edge is represented by a spring model. The nodes connected at the edge are arranged close to each other, the nodes not connected at the edge are arranged far from each other, and all the nodes included in the partial network structure are arranged stably. Calculate each two-dimensional plane coordinate
Information presentation method.   The information presentation program for functioning a computer as each part which comprises the information presentation apparatus of any one of Claims 1-4.