JP4894580B2 - Seasonal analysis system, seasonality analysis method, and seasonality analysis program - Google Patents

Description

  The present invention relates to a seasonal analysis system, a seasonal analysis method, and a seasonal analysis program for analyzing the seasonality of articles posted on the Internet.

  There are a huge number of articles on the Internet. In the case of articles that are updated many times a day (for example, blogs), it takes an enormous amount of time to browse all articles and grasp the latest information. Therefore, it is difficult to understand what themes are now and how exciting they are. Furthermore, the genres (sports, economy, etc.) of many articles posted on the Internet are diverse. Although it is the same genre, many articles written from various viewpoints are published. Therefore, it is difficult to select and browse only necessary articles from a large number of articles posted on the Internet.

  Therefore, there is a demand for a technology that can know what themes are on the Internet and how seasonal the themes are. In addition, such technology is required to be able to accurately grasp Internet themes and seasons. Season means not a sudden rise or a long-term trend due to spam, but a short-term rise due to the popularity of the article.

  In relation to the above, Patent Documents 1 to 3 describe techniques for determining what theme is present depending on the appearance frequency of words. Patent Literature 1 collects articles from a plurality of news sites, extracts words, extracts words with high appearance frequency as important words, sets words related to the important words as related words, and related words for the important words. It is described that the degree of relevance is calculated and displayed. Patent Document 2 describes that, based on the creation time of an article, the newer the creation time of an article, the more weight is given to the words that appear in the article, and the ranking of new words by category is performed. . Patent Document 3 describes that, in a blog for which an update notification has been made, the frequency of occurrence of keywords is counted and a trend keyword in the near future is predicted using the number of appearance frequencies.

  Moreover, the technique which evaluates the popularity and importance with respect to an article in relation to the season is described in Patent Documents 4 to 6. Patent Document 4 describes that the popularity of a corresponding web page is calculated using the number of links from another web page (article) to the corresponding web page. Patent Document 5 describes that the number of homepage accesses is regarded as popular. Patent Document 6 describes that, when calculating the page importance ranking, the links of the web pages in the web page data are aggregated at a certain level and the aggregated links are used.

  Patent Document 7 describes a document classification device that classifies a plurality of articles according to the contents of the articles. Patent Document 7 describes that document feature vectors are generated by counting the number of word appearances for each document data, and a document classification is performed by applying a statistical method to the document feature vectors.

JP 2002-108937 A JP-A-2005-135311 JP 2006-227965 A Japanese Patent No. 3802913 JP 2002-132976 A JP 2006-127529 A JP 2001-101227 A

  An object of the present invention is to provide a technique capable of accurately knowing what themes are on the Internet.

  Another object of the present invention is to provide a technique capable of accurately grasping the seasonality of a theme in an article on the Internet.

  [Means for Solving the Problems] will be described below using the numbers and symbols used in [Best Mode for Carrying Out the Invention]. These numbers and symbols are added in parentheses in order to clarify the correspondence between the description of [Claims] and [Best Mode for Carrying Out the Invention]. However, these numbers and symbols should not be used for the interpretation of the technical scope of the invention described in [Claims].

  The seasonal analysis system of the present invention collects article contents for a plurality of starting articles and a group of articles that can be traced from each of the plurality of starting articles among a plurality of articles posted on the Internet. The article collection unit (1) that records in the article recording unit (2) in association with the article information that identifies the article, and for each starting article collected by the article collection unit (1), the starting article and the starting article A cluster generation unit (3) for generating a cluster information in association with the article information and recording the cluster information in the cluster recording unit (4). Referring to the article recording unit (2) and the cluster recording unit (4), the similarity between different clusters is calculated based on the article content of the articles included in the cluster, and similar based on the calculation result. The Refers to the cluster reconfiguration unit (5) that merges the clusters and generates post-merge cluster information, and the post-merge cluster information and article recording unit (2), and measures the seasonality for each cluster. A part (7), and an output unit (8) for causing the output device to output the result measured by the seasonality measuring unit (7).

  According to the present invention, the cluster is reconfigured by the cluster restructuring unit (5) so that the clusters having similar article contents are gathered together, so that many articles on the Internet can be classified by theme. Thereby, it is possible to accurately grasp what theme is on the Internet. In addition, since the seasonality measuring unit (7) measures the seasonality for each cluster merged by theme, the seasonality for the theme can be grasped.

  In the seasonal analysis system, the cluster reconstruction unit (5) analyzes the content of the article recorded in the article recording unit (2), and generates appearance frequency data in which the word and the appearance frequency are associated with each other. Referring to the article analysis unit (9) to be recorded in the appearance frequency recording unit (10), the cluster recording unit (4), and the appearance frequency recording unit (10), the word and the appearance frequency are associated with each cluster. A feature vector generation unit (11) that generates a cluster feature vector and records it in the feature vector recording unit (12), and calculates a similarity between different clusters based on the cluster feature vector, and similar based on the calculation result It is preferable to include a similarity determination unit (13) that merges clusters to generate cluster information after merging.

  Well, at this time, the article analysis unit (2) preferably refers to the thesaurus dictionary and generates the appearance frequency data so that similar words in the appearance frequency data are merged.

（但し、→ｖiはクラスタiの特徴ベクトルを示し、→ｖjはクラスタjの特徴ベクトルを示す） In the seasonal analysis system described above, the similarity determination unit (14) preferably calculates the similarity according to the following Equation 1 when calculating the similarity between different clusters.

(Note that → vi indicates the feature vector of cluster i, and → vj indicates the feature vector of cluster j)

（但し、α、β、γは定数、ｔはトラックバックの張られてからの経過時間を示す） In the seasonal analysis system described above, the article collection unit (1) further collects and records the article content, and further includes trackback information for identifying the original article of the trackback that is attached to the collected article, It is preferable to collect the stretched time and record it in association with the article information. Here, the seasonality measurement unit (7) calculates the freshness for each cluster and records the freshness calculation unit (15) in the freshness recording unit (16). And a seasonal calculation unit (21). Further, the freshness degree calculation unit (16) calculates the freshness degree based on the number of trackbacks included in the cluster and the time when the trackback is applied. Moreover, it is more preferable that the freshness calculation unit (15) calculates the freshness F (t) by the following formula 2.

(However, α, β, and γ are constants, and t is the elapsed time since the trackback was applied.)

（但し、αはトラックバック元の記事の重みを示し、ｄはトラックバックの前記起点記事からの深さを示す） In the above seasonal analysis system, the article collection unit (1) further collects trackback information for identifying the original article of the trackback that is attached to the collected article when collecting and recording the article content. It is preferable to record in association with article information. Here, the seasonality measurement unit (7) calculates the ripple degree for each cluster and records the ripple degree calculation unit (19) in the ripple degree recording unit (20), and determines the seasonality based on the result of the ripple degree. The seasonal calculation part (21) to calculate is provided. The spillover degree calculation unit (19) calculates the depth from the trackback starting article based on the trackback information, and determines the spillover based on the number of trackbacks included in the cluster and the depth from the trackback starting article. Calculate the degree. At this time, it is more preferable that the ripple degree calculation unit (19) calculates the ripple degree I (d) according to the following formula 3.

(However, α indicates the weight of the article of the trackback source, and d indicates the depth of the trackback from the starting article)

  In the seasonal analysis system described above, the article collection unit (1) further collects evaluation information indicating the reader's evaluation on the collected article and associates it with the article information when collecting and recording the article content. It is preferable to record. Here, the seasonality measurement unit (7) calculates the popularity for each cluster and records the popularity in the popularity recording unit (18), and the seasonality based on the result of popularity. And a seasonal calculation unit (21) for calculating. The popularity calculation unit (21) calculates the popularity based on the number of articles included in the cluster and the evaluation information of the articles. At this time, the evaluation information is more preferably the number of social bookmarks or the number of accesses to the collected articles.

（但し、α、β、γは定数、ｔはトラックバックの張られてからの経過時間を示す）
波及度計算部（１９）は、下記式５により、波及度Ｉ（ｄ）を計算し、

（但し、αはトラックバック元の記事の重みを示し、ｄはトラックバックの前記起点記事からの深さを示す）
前記人気度計算部（１７）は、収集される記事に対するソーシャルブックマーク数又はアクセス数を評価情報として、クラスタ毎の総ソーシャルブックマーク数又は総アクセス数を、人気度Ｐ（ｎ）として計算し、
旬度計算部（２１）は、下記式６により、旬度を計算することが更に好ましい。

（但し、Ｉi(d)はトラックバックiの波及度を示し、Ｆi(t)はトラックバックiの新鮮度を示し、Ｐj(n)は、記事ｊの人気度を示す） In the seasonal analysis system described above, the article collection unit (1) further collects and records an article, and further includes trackback information for identifying an original article of the trackback that is attached to the collected article, It is preferable that collected time and evaluation information indicating the reader's evaluation on the collected articles are collected and recorded in association with the article information. Here, the seasonality measuring unit (7) calculates, for each cluster, the popularity based on the number of articles included in the cluster and the evaluation information of the articles, and records the popularity in the popularity recording unit (18). And calculating the depth from the trackback starting article for each cluster based on the trackback information, the number of trackbacks included in the cluster, and the depth from the trackback starting article A ripple degree calculation unit (19) that calculates a ripple degree based on the data and records it in the ripple degree recording unit (20), and for each cluster, the number of trackbacks included in the cluster and the time at which the trackback was applied Based on the results of the freshness calculation unit (15) that calculates the freshness and records the freshness in the freshness recording unit (16), and the popularity, the spread, and the freshness result are calculated. Prepared with seasonal calculation part (7) That. At this time, the freshness calculation unit (15) calculates the freshness F (t) according to the following equation 4,

(However, α, β, and γ are constants, and t is the elapsed time since the trackback was applied.)
The spillover degree calculation unit (19) calculates the spillover degree I (d) by the following formula 5,

(However, α indicates the weight of the article of the trackback source, and d indicates the depth of the trackback from the starting article)
The popularity calculating unit (17) calculates the total number of social bookmarks or the total number of accesses for each cluster as the popularity P (n) using the number of social bookmarks or the number of accesses to the collected articles as evaluation information,
More preferably, the seasonality calculation unit (21) calculates the seasonality according to the following Equation 6.

(Where Ii (d) indicates the influence of trackback i, Fi (t) indicates the freshness of trackback i, and Pj (n) indicates the popularity of article j)

  ADVANTAGE OF THE INVENTION According to this invention, the technique which can know exactly what theme exists on the internet is provided.

  According to the present invention, there is further provided a technology capable of accurately grasping the seasonality of a theme in an article on the Internet.

(First embodiment)
Hereinafter, the first embodiment will be described with reference to the drawings. The seasonal analysis system of the present embodiment automatically extracts articles with high seasonality from a large number of articles on the Internet, and displays information for identifying the articles. The articles to be analyzed include articles included in a user participation type web system group with a link support system, such as a blog or WIKI.

  FIG. 1 is a schematic block diagram illustrating a configuration of a seasonality analysis system according to the present embodiment. The seasonal analysis system is stored in a ROM (Read only memory) or the like, and is executed from a seasonal analysis program executed by the CPU and a storage device (RAM, hard disk, etc.) that records the processing contents and results of the seasonal analysis program. Composed. Specifically, as shown in FIG. 1, an article collection unit 1, an article recording unit 2, a cluster generation unit 3, a cluster recording unit 4, a cluster reconfiguration unit 5, a merged cluster recording unit 6, a seasonality measuring unit 7 and a display unit 8.

  The article collection unit 1 implements a function of periodically acquiring articles from the Internet, generating article data, and storing the article data in the article recording unit 2. Here, the article collection unit 1 stores only updated information and new information in the article recording unit 2.

  The cluster generation unit 3 sets a cluster for at least one article collected by the article collection unit 1 and stores it in the cluster recording unit 4 as cluster information.

  The cluster reconfiguration unit 5 refers to the article recording unit 2 and the cluster recording unit 4 to determine the similarity between clusters, generates post-merge cluster information by merging similar clusters, and records post-merge cluster information This is stored in the unit 6. The cluster reconstruction unit 5 includes an article analysis unit 9, an appearance frequency recording unit 10, a feature vector generation unit 11, a feature vector recording unit 12, a similarity determination unit 11, and a similarity recording unit 14.

  The seasonality measuring unit 7 measures the seasonality for each cluster after merging with reference to the article recording unit 2 and the merged cluster recording unit 6. The seasonality measuring unit 7 includes a freshness calculating unit 15, a freshness recording unit 16, a popularity calculating unit 17, a popularity recording unit 18, a ripple calculating unit 19, a ripple recording unit 20, a seasonality calculating unit 21, and A seasonal recording unit 22 is provided.

  The display unit 8 displays the measurement result of the seasonality measurement unit 7 on the display screen.

  This seasonal analysis system operates as described below to realize its functions. FIG. 2 is a flowchart showing a seasonality analysis method according to this embodiment. This seasonal analysis method includes a step of collecting articles (S001), a step of clustering (S002), a step of determining similarity to reconstruct a cluster (S003), a step of measuring seasonality (S004), And a displaying step (S0005). The operation in each step will be described in detail below.

((Step S001; Collecting articles))
FIG. 3 is a flowchart showing an operation when the article collection unit 1 collects articles posted on the Internet. In the description of this step, a case where a plurality of articles (A to H) are posted on the Internet as shown in FIG. 4 will be described as an example. In FIG. 4, an arrow indicates a link by trackback, and for example, article A is tracked back from article B, and a link is provided from article A to article B.

Step S101
First, the article collection unit 1 selects an article as a starting point (hereinafter referred to as a starting article). Here, it is assumed that article A is selected as the starting article. As the starting article, for example, an article with a large number of registered social bookmarks, an article with a large number of accesses, an article specified by the user, or the like may be selected. The article collection unit 1 accesses the selected starting article A, collects data related to the article A (article data), and stores it in the article recording unit 2.

  FIG. 5 is a conceptual diagram showing data recorded in the article recording unit 2. Of the items shown in this figure, the article collection unit 2 includes article contents (title and body text), URL (article contents), creation time of article A, trackback source URL for article A, and trackback links. Time and evaluation information (in this embodiment, the number of social bookmarks) are collected. These are recorded in the article recording unit 2 in association with the article ID (article information; A) for specifying the article. At this time, the item “visited” shown in FIG. 5 is set to NO. Then, the process proceeds to the next step (S102). In FIG. 5, items relating to articles B, C, and D are filled in, and the item “visited” is also YES, but FIG. 5 shows an example of the final state. Yes, it does not show the state after the processing in this step.

Step S102
Next, the article collection unit 2 refers to the article recording unit and determines whether there is an unvisited article. If it exists, the process proceeds to the next step S103, and if it does not exist, the process related to article collection ends. At this stage, since the visit of the starting article A is not visited, the process proceeds to the next step S103.

Step S103, S104
Next, the article collection unit 1 selects the youngest article with the article ID from unvisited articles. Here, the starting article A is selected (S103). Then, it is confirmed whether or not a trackback is stretched on the selected article. If the trackback is stretched, the process proceeds to step S105, and if not, the process proceeds to S106. Here, since the trackback from the article B and the article C is applied to the selected starting article A, the process proceeds to S105 (S104).

Step S105
Next, the article collection unit 1 accesses the trackback source article of the article selected in S103, collects data related to the article, and stores it in the article recording unit 2 as in the process of S101. Here, article B and article C are accessed and collected. At this time, article B and article C are not visited. Thereafter, the process proceeds to next Step S106.

Step S106
Next, the article collection unit 1 records in the article recording unit 2 that the article (article A) selected in S103 has been visited, and returns to the process of S102. In S102, since article B and article C have not been visited, the process proceeds to S103, and article B having a lower article ID is selected. In S104, since the article B has no trackback source, the process proceeds to S106, and the article recording unit 2 records that the article B has been visited. Such processing is repeated, and data relating to articles is collected and recorded in the article recording unit 2 for all article groups that can be accessed by following links from the starting article A. Similar processing is repeated for the other starting articles (D and G), and data relating to the articles is collected and recorded.

  In the above description of the series of processes, the case has been described in which the article contents are collected for all articles that can be traced from the starting article. However, for example, in the second and subsequent processes, the article recording unit 2 has already been collected. When article contents are stored, only updated articles and new articles need to be collected.

((Step S002)); Clustering Processing Next, the cluster generation unit 3 collects a plurality of articles collected by the article recording unit 21 from a starting article and a set of articles that can follow a trackback link from the starting article. Set as a cluster. The cluster generation unit 3 allocates a unique ID (hereinafter referred to as a cluster ID) to the set cluster, associates an article ID in the cluster, generates cluster information, and records it in the cluster recording unit 4.

  6 and 7 are conceptual diagrams showing an example of cluster information recorded in the cluster recording unit 4, and correspond to the example given in FIG. In this example, a set composed of a starting article A and articles B and C that can follow links from the starting article A is set as a cluster C1. Similarly, a set of starting articles D, articles E, and F is set as the cluster C2, and a set of starting articles G and articles H is set as the cluster C3. In FIG. 6, article names A and B are shown for easy understanding of the explanation, but in reality, article IDs (B1, B2,...) Correspond to cluster IDs indicating clusters. It is attached. In the following description, article names (A, B, C,...) May be described instead of article IDs (B1, B2,...).

  The set of articles that can follow the trackback link can be grasped by referring to the URL of the trackback source recorded in the article recording unit 2, for example.

((Step S003)); Determination of Similarity, Cluster Reconfiguration Next, FIG. 8 is a flowchart showing the operation in this step. This operation is realized by the operation of the cluster reconfiguration unit 5.

Step S201, 202; Article Selection First, the article analysis unit 9 acquires all the article contents recorded in the article recording unit 2 (S201). Then, morphological analysis is performed on the acquired article. Thereby, the content of the article is decomposed into words, and nouns are extracted from the words (S202). The purpose of this process is to determine the similarity of article themes (themes), and there is no need for parts of speech that are not suitable for expressing the themes. Accordingly, only nouns are extracted from the result of the morphological analysis process, excluding punctuation marks, verbs, adjectives and the like. The article analysis unit 9 counts the number of appearances of nouns appearing in the article, gives unique IDs (W1, W2,...) For each noun, generates appearance frequency data exemplified in FIG. Record in the appearance frequency recording unit 10.

Step S203: Generation of Feature Vector Subsequently, the feature vector generation unit 11 refers to the appearance frequency data and the cluster information, and totals the number of appearances of each word for each cluster. Then, as shown in FIG. 10, for each cluster, a feature vector in which the word ID (W1, W2,...) Is associated with the number of appearances is generated and recorded in the feature vector recording unit 12. At this time, in all the clusters (is it correct in all the clusters?), A feature vector is generated by selecting an upper word (for example, upper 10 words) in descending order of appearance frequency. Note that the example of FIG. 10 does not correspond to the example shown in FIG. 8 for convenience of explanation.

（但し、→ｖiはクラスタiの特徴ベクトルを示し、→ｖjはクラスタjの特徴ベクトルを示す） Steps S204 and 205: Calculation and Comparison of Similarity Subsequently, the similarity determination unit 13 acquires feature vectors of two different clusters from the feature vector recording unit 12. Then, the similarity is calculated from the angle formed by the two acquired feature vectors. In this embodiment, the cosine of the angle formed by the feature vectors of the two clusters (i and j) is defined by the similarity as shown by the following equation (7).

(Note that → vi indicates the feature vector of cluster i, and → vj indicates the feature vector of cluster j)

Taking the case of FIG. 10 as a specific example, the similarity calculation method will be described more specifically. Here, the similarity between cluster 1 and cluster 2 is calculated. When the similarity between the clusters 1 and 2 is calculated by the above equation 7, the calculation is performed as follows, and a result of similarity = 0.01 is obtained.
Similarity = (3 × 2 + 0 × 2 + 9 × 0 + 7 × 0) / {√ (3 ² +0 ² +0 ² +9 ² +7 ² ) × √ (2 ² +3 ² +2 ² +0 ² +0 ² ) = 0.01

  The similarity determination unit 13 determines whether the two clusters are similar from the calculated similarity. The more words that appear in common in the feature vectors of two different clusters and the more words that appear closer to each other, the closer the angle formed by the feature vectors is to 1, and the higher the similarity. Therefore, a threshold value of similarity is set in advance, and the feature vector angle is closer than the threshold value (similarity is closer to 1 than the threshold value), the clusters are determined to be similar, and the same theme is handled. Is determined. In the above-described example, if the threshold value is set to 0.07, similarity = 0.01, which is not closer to 1 than the threshold value, so it is determined that they are not similar. (Step S205). If they are similar, the operation of the next step S206 is performed, and if they are not similar, the operation of S207 is performed.

  The similarity calculated in the above-described processing (S204, 205) is stored in the similarity recording unit 14 as shown in FIG.

Step S206; Merging Similar Clusters When the two clusters are similar in the process of S205, the similarity determination unit 13 refers to the cluster information, adds up the number of appearances of each word, and merges the similar clusters. Thereby, a new feature vector is generated. In addition, the similarity determination unit refers to the cluster information, generates post-merge cluster information to which a cluster in which similar clusters are merged is added, and stores the merged cluster information in the post-merge cluster recording unit. FIG. 12 is a conceptual diagram showing an example of merged cluster information. In the example of FIG. 12, the cluster C1 and the cluster C2 are similar to each other, and an example in which a cluster C4 including the clusters 1 and 2 is added is shown. For convenience of explanation, it does not correspond to the above-described figure. The clusters in the merged feature vector are set so as not to be subjected to similarity comparison, and the process of the next step S207 is performed.

Step S207: End Determination The above-described series of processing is repeated for all combinations of clusters, and when the similarity determination is completed for all combinations, the processing ends. Since the merged cluster information reconstructed in this way is a collection of a plurality of articles based on the article content, it can be regarded as a collection of articles by theme. That is, articles in different clusters can be regarded as articles having different themes, and articles in the same cluster can be regarded as articles of the same theme. In addition, it can be said that the cluster feature vector merged in S206 indicates a keyword frequently used in the cluster. Therefore, if a keyword is displayed for each cluster, the user can grasp what theme exists.

((Step S004)); Measurement of Seasonal Value Subsequently, the seasonality measurement unit 7 measures the seasonality. FIG. 13 is a flowchart showing an operation related to seasonal measurement. The seasonality measurement unit 7 calculates freshness (S301), popularity (S302), and spread (S303), and calculates the seasonality for each cluster from these values (S304). The calculation of freshness, popularity, and spread may be performed in any order. In addition, when there are merged clusters of multiple clusters (merge clusters), the popularity degree, freshness degree, and ripple degree are calculated for each component cluster of the merge cluster at the time of seasonal calculation. The values are added together to obtain the merge cluster evaluation value. Details of the operation in each step will be described below.

Step S301: Calculation of Freshness The fresh time calculation unit 15 refers to the article recording unit 15, and describes the article ID, the trackback source article information (trackback source URL), and the time when the trackback is set. And get. Further, the current time is acquired from a timer function unit (not shown). Information for identifying the trackback source article and the trackback destination article by calculating the elapsed time (difference of the trackback extension time from the current time) for the one trackback. ID) is held in association with each other (FIG. 14). In FIG. 14, article names (A, B,...) Are described instead of article IDs for easy understanding. The elapsed time is calculated for all trackbacks recorded in the article recording unit 15. For each trackback, the freshness is calculated based on the elapsed time.

（但し、α、β、γは定数、ｔはトラックバックの張られてからの経過時間を示す）
尚、式中、ｔは経過時間を示し、ｔ＞＝０である。αは、新鮮度の減少量の幅に関する定数である。βは、新鮮度の傾きを示す。γは、新鮮度の初期値を示す値である。 In this embodiment, the freshness of the trackback is set to decrease exponentially as shown in FIG. 15 with respect to the elapse of time after the trackback is stretched. A case where the degree F (t) is obtained will be described as an example.

(However, α, β, and γ are constants, and t is the elapsed time since the trackback was applied.)
In the formula, t indicates elapsed time, and t> = 0. α is a constant related to the range of the amount of decrease in freshness. β indicates the slope of freshness. γ is a value indicating an initial value of freshness.

  The more new trackbacks, the more attention is likely to be placed on the articles behind them. Conversely, if there is only an old trackback, it will be low. However, a certain value remains even after a certain period of time. Therefore, it is preferable to use an exponential function that decreases with the passage of time, as shown in Equation 8 above.

  As a specific example, it is assumed here that α = 10, β = 10, and γ = 0, and the elapsed time of each trackback is calculated as shown in FIG. At this time, the freshness calculation unit 15 calculates according to Equation 9, and the freshness of the trackback between A and B = 1.4, the freshness of the trackback between C and A = 3.7, and the trackback between D and B The calculation result is freshness = 6.1 (however, in this example, the third decimal place is rounded off).

  The freshness calculation unit 15 records the calculated freshness of the trackback in the freshness recording unit 16 in association with the information for specifying the trackback (see FIG. 16; however, an example is shown. Does not correspond to the contents of the figure).

  Above, the process which concerns on calculation of freshness is complete | finished.

  If a writer of an article B evaluates another article A at a time close to the current time, evaluates the blog and puts a trackback, the evaluation value of the freshness of the article A at the trackback destination is considered to increase. Therefore, the trackback freshness calculated above can be regarded as the freshness of the trackback destination article. Therefore, the freshness level can also be used as a parameter of article excitement. Further, when the freshness of the cluster is regarded as being proportional to the sum of the freshness of the trackback, the freshness of the cluster can be obtained by adding the freshness of the trackback in the cluster. Therefore, it is also useful that the freshness calculation unit 8 calculates these as necessary and displays them on the display screen by the display unit 8.

Step S302: Popularity Calculation The popularity calculation unit 17 refers to the article recording unit 2 and calculates the popularity of the article based on the evaluation information of the article (the number of social bookmarks in this embodiment). As shown in FIG. 17, if the number of social bookmarks of article A is 10, article B is 5, article C is 3, and article D is 2, the popularity of article A is 10, and the popularity of article B is 5. The popularity of article C is 3, and the popularity of article D is 2. The popularity calculated in this way is recorded in the popularity recording unit 114. FIG. 19 is an example of popularity degree data in the first embodiment of the present invention. The popularity calculating unit 17 records the popularity in association with the article ID in the popularity recording unit 118 (see FIG. 18). The evaluation information is not limited to the number of social bookmarks as long as it is information indicating the evaluation on the reader side of the article, and other data such as the number of accesses may be used.

  It should be noted that the popularity of articles in a cluster may be totaled to be the popularity of the cluster.

Step S303: Calculation of Ripple Degree The ripple degree calculator 19 refers to the article recording unit 2 and calculates the ripple degree based on the path length (depth) from the starting article of the track back. Here, it is assumed that the path length is the number of links (trackbacks) existing between the starting article and the trackback destination article. The more distant articles (the longer the path length) are from the starting article, and the more links (trackbacks), the more the theme is considered to be exciting. Therefore, the longer the path length, the higher the ripple degree. For concrete explanation, as shown in FIG. 19, a trackback is stretched from the articles B and C to the starting article A, and a trackback is stretched from the article D to the article B. And At this time, the popularity calculation unit 17 calculates that the track length of the trackback from the article B to A and the trackback from the article C to A is 1. Also, it is calculated that the trackback path length for articles D to B is 2.

尚、上式中において、ｄはパス長である。αは、リンクのパス長への重みである。この重みは、信頼性の高い記事（例示；アクセス数の多い記事）から直接リンクされている場合などには大きくし、信頼性の低い記事から直接リンクされている場合などは値を小さくする様にすればよい。 Further, the spread degree may be calculated by weighting in consideration of the reliability of the article. Thus, as an example when the reliability of the article is taken into consideration, the ripple degree I (d) can be defined as in the following Expression 9.

In the above formula, d is a path length. α is a weight to the path length of the link. This weight is increased when linked directly from highly reliable articles (example: articles with a large number of accesses), and is decreased when linked directly from unreliable articles. You can do it.

  The ripple degree calculation unit 19 records the ripple degree calculated as described above in the ripple degree recording unit 20 in association with the trackback source article ID and the trackback destination article ID (see FIG. 20). As described in the freshness level, the trackback ripple may be taken as the ripple of the trackback destination article.

Step S304: Measurement of seasonality The seasonality calculation unit 21 calculates the seasonality based on the ripple, freshness, and popularity calculated in the processes of S301 to 303. At this time, the seasonality is calculated for each cluster based on the merged cluster information.

（但し、Ｉi(d)はトラックバックiの波及度を示し、Ｆi(t)はトラックバックiの新鮮度を示し、Ｐj(n)は、記事ｊの人気度を示す） The following equation 10 is an equation showing an example of seasonal calculation.

(Where Ii (d) indicates the influence of trackback i, Fi (t) indicates the freshness of trackback i, and Pj (n) indicates the popularity of article j)

More specifically, it is assumed that articles included in the cluster C1 are linked by trackback as shown in FIG. It is assumed that the result shown in FIG. 16 as the freshness, the result shown in FIG. 20 as the spread, and the result shown in FIG. 18 as the popularity are obtained. In this case, the seasonality of the theme of cluster C1 is
Season of the theme of the blog cluster C1 = Ripple of trackback between (A and B) × Freshness of trackback between (A−B) + Ripple of trackback between (A−C) × (A−C) Trackback Freshness + (BD) Trackback Ripple x BD Trackback Freshness + Sum of Popularity of Blog Articles A to D = 1 × 1.4 + 1 × 3.7 + 2 × 6.1 + (10 + 5 + 3 + 2)
= 37.3
It becomes.

  The seasonality calculation unit 21 records the seasonality of the theme calculated in this way in the seasonality recording unit 22 in association with the cluster ID (see FIG. 21; however, the clusters C2 and C3 are described above). Does not correspond to the example).

  As described above, if the trackback ripple is regarded as the ripple of the trackback destination article and the trackback freshness is regarded as the freshness of the trackback destination article, the seasonality can be calculated for each article. Specifically, it may be obtained as “season of article” = spreading degree × freshness + popularity.

((Step S005; Display))
The display unit 12 refers to the article recording unit 2 and the merged cluster information, and acquires the title of the starting article for each cluster. In addition, referring to the seasonal recording unit 22, the clusters are rearranged in descending order of the seasonality, and as shown in FIG. 22, the seasonality corresponding to the title of the starting article is displayed on the display screen. At this time, the keyword of the cluster is displayed with reference to the cluster feature vector (see FIG. 10). Further, when displaying the merge cluster, it may be displayed hierarchically. Further, articles other than the starting article may be displayed as related articles. At this time, if the seasonality of each article is obtained, the seasonality of each article may be displayed together.

  As described above, according to the present embodiment, a plurality of articles on the Internet are clustered based on the article content, and the title of the starting article, keywords, etc. are displayed in order from the highest season, so the user can You can find out what themes exist as recent seasonal themes without browsing all the articles. In addition, it is possible to grasp the excitement of the theme.

  In addition, since the ripple degree is used to calculate the seasonality, it is possible to reflect in the evaluation of whether it is seasonal how much the theme is attracting attention.

  In addition, since freshness is used in calculating seasonality, old evaluations and recent evaluations are not treated equivalently, and more recent evaluations can be weighted and reflected in seasonality.

(Second Embodiment)
A second embodiment of the present invention will be described. FIG. 23 is a block diagram schematically showing the configuration of the seasonality analysis system of this embodiment, and FIG. 24 is a flowchart showing the operation method. The seasonal analysis system of the present embodiment is different from the first embodiment in that a thesaurus analysis unit 23 is added to the article analysis unit 9 (FIG. 23), and the operation method includes a step of performing a thesaurus analysis of words. (S202A) is added. Other configurations and operations will be described with the same reference numerals.

  When the article analysis unit 9 generates appearance frequency data (S202), the thesaurus analysis unit 23 performs a thesaurus analysis on the generated appearance frequency data with reference to a thesaurus dictionary (not shown) (S202A). That is, it is determined whether or not there is a thesaurus (similar) word among the extracted words, and the appearance frequencies of the words determined as the thesaurus are merged and handled as one word. The thesaurus analysis unit 9 records the appearance frequency data in which similar words are merged in this manner in the appearance frequency recording unit 10 as post-merge appearance frequency data (AS201 in FIG. 24). In the processing after step S203, the appearance frequency data after merging is used instead of the appearance frequency data in the first embodiment.

  According to the present embodiment, not only the appearance frequency of words but also the thesaurus of the word is analyzed and the similar words are combined into one, so that the variation of the words of the articles can be suppressed, and the similarity of the clusters The accuracy at the time of determination can be improved.

  When the thesaurus analysis unit 23 performs the thesaurus analysis, the thesaurus analysis unit 23 may perform the thesaurus analysis by storing the thesaurus dictionary as electronic data, or the thesaurus using the thesaurus dictionary on the Internet. Analysis may be performed.

It is a schematic block diagram of the seasonal analysis system which concerns on 1st Embodiment. It is a flowchart which shows the seasonal analysis method which concerns on 1st Embodiment. It is a flowchart which shows operation | movement of an article collection part. It is explanatory drawing which shows the example of the coupling | bonding of the trackback of an article. It is a conceptual diagram showing the data content recorded on an article recording part. It is a conceptual diagram showing the data content recorded on a cluster recording part. It is explanatory drawing for demonstrating operation | movement of a cluster production | generation part. It is a flowchart which shows operation | movement of a cluster reconstruction part. It is a conceptual diagram which shows appearance frequency data. It is a conceptual diagram which shows a cluster feature vector. It is a conceptual diagram showing the data content recorded on a similarity recording part. It is explanatory drawing for demonstrating cluster information after a merge. It is a flowchart which shows operation | movement of a seasonal measurement part. It is a conceptual diagram which shows the relationship between a track back and track back time. It is explanatory drawing for demonstrating the relationship between freshness and elapsed time. It is a conceptual diagram showing the data content recorded on a freshness recording part. It is explanatory drawing for demonstrating the relationship between an article and evaluation information. It is a conceptual diagram showing the data content recorded on a popularity recording part. It is explanatory drawing for demonstrating the relationship between a track back and path length. It is a conceptual diagram showing the data content recorded on a ripple recording part. It is a conceptual diagram showing the data content recorded on a seasonal recording part. It is an example figure showing the content displayed on a display screen. It is a schematic block diagram which shows the structure of the seasonal analysis system of 2nd Embodiment. It is a flowchart which shows operation | movement of the cluster reconstruction part of the seasonal analysis method of 2nd Embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Article collection part 2 Article recording part 3 Cluster production | generation part 4 Cluster recording part 5 Cluster reconfiguration | reconstruction part 6 Post-merge cluster recording part 7 Seasonal measurement part 8 Display part 9 Article analysis part 10 Appearance frequency recording part 11 Feature vector generation part 12 Feature vector recording unit 13 Similarity determining unit 14 Similarity recording unit 15 Freshness calculating unit 16 Freshness recording unit 17 Popularity calculating unit 18 Popularity recording unit 19 Ripple calculating unit 20 Ripple recording unit 21 Seasonal calculating unit 22 Seasonal recording part 23 Thesaurus analysis part

Claims (25)

Select a plurality of starting articles from a plurality of articles posted on the Internet, collect article contents for each of the plurality of starting articles and a group of articles that can be traced from each starting article by a link, and An article collection unit that generates article data associated with the article information to be identified;
For each starting article, a cluster generation unit that sets the starting article and a set of articles that can be traced from the starting article as a cluster, and generates cluster information that associates the cluster with the article information. When,
Based on the article data and the cluster information, the similarity between different clusters is determined based on the article contents of the articles included in the cluster, and the similar clusters are merged by merging the similar clusters based on the determination result. A cluster reconfiguration unit that generates information;
A seasonality measuring unit that refers to the cluster information after merging and the article data, and measures the seasonality for each cluster,
An output unit for causing the output device to output the result measured by the seasonality measuring unit;
Seasonal analysis system. The seasonal analysis system according to claim 1,
The cluster reconfiguration unit
Analyzing the article content of the article based on the article data, and generating an appearance frequency data in which words and appearance frequencies are associated with each other,
Based on the cluster information and the appearance frequency data, for each cluster, a feature vector generation unit that generates a cluster feature vector in which a word and an appearance frequency are associated;
A similarity determination unit that calculates similarity between different clusters based on the cluster feature vector, merges similar clusters based on a determination result, and generates the cluster information after merging. Analysis system. A seasonal analysis system according to claim 2,
Furthermore,
Thesaurus analysis is performed on the appearance frequency data generated by the article analysis unit, and a thesaurus analysis unit that generates merged appearance frequency data obtained by merging similar words,
The feature vector generation unit generates the cluster feature vector with reference to the merged appearance frequency data. A seasonality analysis system according to any one of claims 1 to 3,
The similarity determination unit calculates the similarity between cluster i and cluster j according to the following formula 1.

(Note that → vi indicates the feature vector of cluster i, and → vj indicates the feature vector of cluster j)
Seasonal analysis system. The seasonal analysis system according to claim 1,
In generating the article data, the article collection unit further includes trackback information for identifying an original article of a link attached to a reference-side article by a reference side with respect to an article to be collected, and a time at which the trackback is applied. And the article data in association with the article information,
The seasonality measuring unit is
A freshness calculator for calculating freshness;
Based on the result of the freshness, with a seasonality calculation unit that calculates the seasonality for each cluster,
The freshness calculation unit is a seasonal analysis system that calculates the freshness of the trackback based on the trackback information and the time when the trackback is applied. A seasonal analysis system according to claim 5,
The freshness calculation unit calculates the freshness F (t) according to the following equation 2 when calculating the freshness for each cluster.

(However, α, β, and γ are constants, and t is the elapsed time since the trackback was applied.)
Seasonal analysis system The seasonal analysis system according to claim 1,
The article collection unit, when generating the article data, further collects trackback information for specifying an original article of the trackback that is attached to the collected article, and associates with the article information,
The seasonality measuring unit is
A ripple calculation unit for calculating the ripple for each cluster;
Based on the results of the ripples, a seasonal calculation unit for calculating the seasonality,
The ripple calculation unit calculates the depth of the trackback from the starting article based on the trackback information, and determines the number of the trackbacks included in the cluster, the depth of the trackback from the starting article, A seasonal analysis system for calculating the spread based on the above. The seasonal analysis system according to claim 7,
The ripple degree calculation unit calculates the ripple degree I (d) according to Equation 3 below.

(However, α indicates the weight of the article of the trackback source, and d indicates the depth of the trackback from the starting article)
Seasonal analysis system. The seasonal analysis system according to claim 1,
In generating the article data, the article collection unit further collects evaluation information indicating a reader's evaluation on the collected articles, and associates with the article information,
The seasonality measuring unit is
A popularity calculation unit for calculating the popularity for each cluster;
Based on the result of popularity, a seasonal calculation unit for calculating the seasonality,
The popularity calculation unit is a seasonal analysis system that calculates the popularity based on the number of articles included in the cluster and the evaluation information of articles. The seasonal analysis system according to claim 9,
The evaluation information is a seasonal analysis system that is the number of social bookmarks or the number of accesses to collected articles. The seasonal analysis system according to claim 1,
When generating the article data, the article collection unit further includes trackback information for identifying an original article of the trackback that is applied to the collected article, the time when the trackback is applied, and the collected article. Collecting evaluation information indicating the reader's evaluation and associating it with the article information;
The seasonality measuring unit is
For each cluster, a popularity calculating unit that calculates the popularity based on the number of articles included in the cluster and the evaluation information of the articles;
For each cluster, the depth of the trackback from the starting article is calculated based on the trackback information, and based on the number of trackbacks included in the cluster and the depth of the trackback from the starting article. And a ripple degree calculation unit for calculating the ripple degree,
For each cluster, a freshness calculation unit that calculates the freshness based on the number of the trackbacks included in the cluster and the time at which the trackback is stretched;
A seasonality analysis system comprising: a seasonality calculation unit that calculates seasonality based on the popularity, the spread, and the freshness results. The seasonal analysis system according to claim 11,
In calculating the freshness for each cluster, the freshness calculation unit calculates the freshness F (t) according to the following formula 4.

(However, α, β, and γ are constants, and t is the elapsed time since the trackback was applied.)
The ripple degree calculation unit calculates the ripple degree I (d) according to the following formula 5.

(However, α indicates the weight of the article of the trackback source, and d indicates the depth of the trackback from the starting article)
The popularity calculating unit calculates the number of social bookmarks or the number of accesses to the collected article as the evaluation information, calculates the total number of social bookmarks or the total number of accesses for each cluster as the popularity P (n),
The seasonality calculation unit calculates seasonality according to the following formula 6.

(Where Ii (d) indicates the influence of trackback i, Fi (t) indicates the freshness of trackback i, and Pj (n) indicates the popularity of article j)
Seasonal analysis system. Among a plurality of articles posted on the web, article information that collects article contents and identifies articles about a plurality of starting articles and a group of articles that can be traced from each of the plurality of starting articles. An article collection step for generating article data in association with each other;
For each starting article collected in the article collecting step, a set including the starting article and a group of articles that can be traced from the starting article by a link is defined as a cluster, and the correspondence between the cluster and the article information is indicated. A cluster generation step for generating cluster information;
Clusters that calculate similarity between article contents between different clusters based on the article data and the cluster information, merge similar clusters based on the calculation results, and generate cluster information after merging A reconfiguration step;
A seasonality measuring step of measuring seasonality for each cluster based on the cluster information after merging and the content of the articles collected in the article collecting step;
An output step of causing the output device to output the measurement result of the seasonal measurement step;
A seasonal analysis method comprising: The seasonal analysis method according to claim 13,
The cluster reconfiguration step includes:
Analyzing the content of the article collected in the article collecting step, and generating an appearance frequency data generating step for generating appearance frequency data in which words and appearance frequencies are associated with each other,
Based on the cluster information and the appearance frequency data, a cluster feature vector generation step for generating a cluster feature vector in which a word and an appearance frequency are associated for each cluster;
A similarity analysis step comprising: calculating similarity between different clusters based on the cluster feature vector, combining similar clusters based on a determination result, and generating cluster information after merging. . The seasonal analysis method according to claim 14, wherein
A seasonality analysis method for generating the appearance frequency data so that similar words are merged with reference to a thesaurus dictionary in the appearance frequency data generation step. A seasonality analysis method according to any one of claims 13 to 15,
In calculating the similarity between the different clusters in the similarity determination step,
Calculate the similarity using Equation 7 below

(Note that → vi indicates the feature vector of cluster i, and → vj indicates the feature vector of cluster j)
Seasonal analysis method. The seasonal analysis method according to claim 13,
In the article collection step, in generating the article data, the article further collects trackback information for identifying the original article of the trackback applied to the collected article and the time when the trackback was applied, and the article The article-data is associated with the information,
The seasonal measurement step includes
A freshness calculating step for calculating freshness for each cluster;
A seasonality calculating step for calculating the seasonality based on the freshness result,
A seasonality analysis method for calculating the freshness in the freshness calculation step, based on the number of the trackbacks included in the cluster and the time when the trackback is applied. The seasonal analysis system according to claim 17,
In the freshness calculation step, the freshness F (t) is calculated by the following equation 8 when calculating the freshness for each cluster.

(However, α indicates the weight of the article of the trackback source, and d indicates the depth of the trackback from the starting article)
Seasonal analysis method. The seasonal analysis method according to claim 13,
In the article collection step, when generating the article data, further collecting trackback information for specifying an original article of the trackback that is attached to the collected article, and obtaining the article data,
The seasonal measurement step includes
A ripple calculation step for calculating the ripple for each cluster;
A seasonality calculating step for calculating the seasonality based on the result of the spread,
In the ripple degree calculating step, the depth of the trackback from the starting article is calculated based on the trackback information, and the number of the trackbacks included in the cluster, the depth of the trackback from the starting article, A seasonal analysis method for calculating the ripple degree based on the above. The seasonal analysis method according to claim 19, wherein
In the ripple calculation step, the ripple I (d) is calculated by the following formula 9.

(However, α indicates the weight of the article of the trackback source, and d indicates the depth of the trackback from the starting article)
Seasonal analysis method. The seasonal analysis method according to claim 13,
In the article collection step, when generating the article data, further collecting evaluation information indicating a reader's evaluation on the collected articles, and associating with the article information as the article data,
The seasonal measurement step includes
A popularity calculation step for calculating the popularity for each cluster;
A seasonality calculating step for calculating the seasonality based on the result of the popularity, and
A seasonality analysis method for calculating the popularity based on the number of articles included in the cluster and the evaluation information of articles in the popularity calculating step. The seasonal analysis method according to claim 21,
The evaluation information is a seasonal analysis method that is the number of social bookmarks or the number of accesses to collected articles. The seasonal analysis method according to claim 13,
In the article collection step, in generating the article data, the trackback information for identifying the original article of the trackback that is applied to the collected article, the time when the trackback is applied, and the collected article Collect evaluation information indicating the reader's evaluation as the article data,
The seasonal measurement step includes
For each cluster, a popularity calculation step for calculating the popularity based on the number of articles included in the cluster and the evaluation information of the articles;
For each cluster, the depth of the trackback from the starting article is calculated based on the trackback information, and based on the number of trackbacks included in the cluster and the depth of the trackback from the starting article. And a ripple degree calculating step for calculating the ripple degree,
For each cluster, a freshness calculation step of calculating a freshness based on the number of the trackbacks included in the cluster and the time when the trackback is stretched;
A seasonality calculating method for calculating a seasonality based on the results of the popularity degree, the ripple degree, and the freshness degree, and a seasonality analysis method. The seasonal analysis method according to claim 23, wherein
In the freshness calculation step, the freshness F (t) is calculated by the following equation 10;

(However, α, β, and γ are constants, and t is the elapsed time since the trackback was applied.)
In the ripple degree calculation step, the ripple degree I (d) is calculated by the following equation 11;

(Where α indicates the weight of the article of the trackback source, and d indicates the depth of the trackback from the starting article)
The popularity calculating unit calculates the number of social bookmarks or the number of accesses to the collected article as the evaluation information, calculates the total number of social bookmarks or the total number of accesses for each cluster as the popularity P (n),
The seasonality calculation unit calculates the seasonality according to the following formula 12.

(Where Ii (d) indicates the influence of trackback i, Fi (t) indicates the freshness of trackback i, and Pj (n) indicates the popularity of article j)
Seasonal analysis method.   A seasonality analysis program for causing a computer to execute the seasonality analysis method according to any one of claims 13 to 24.

Images