JP2017199113A - Burst detection device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To detect a burst topic in a wide range of users.SOLUTION: A burst detection device 1 includes: a contributed information history storing part 10 for storing a contributed information history relating to the contributed information that is contributed by a user; a topic information storage part 13 for storing a user affiliation probability of a user to the topic and a contributed information affiliation probability of the contributed information to the topic; a weighted appearance probability calculation part 14 for calculating a weighted appearance probability of the topic within a unit time on time series in the contributed information history based on the stored contributed information history, a user affiliation probability, and the contributed information affiliation probability, on the basis of the user affiliation probability of the user contributed the contributed information within the unit time to the topic and the contributed information affiliation probability of the contributed information contributed within the unit time to the topic; and a burst detection part 15 for detecting a burst topic on the time series on the basis of the calculated weighted appearance probability of the topic.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a burst detection device that detects topics bursted in time series among topics (topics) of posted information posted by a user.

  Conventionally, in microblogs where users mainly post their own situations and miscellaneous notes in short sentences (tweets), a technique for detecting bursts, which is a phenomenon in which posts on a topic increase rapidly, There is a known method for detecting a burst of postings including. For example, in Patent Document 1 below, a burst degree indicating the appearance frequency of words in a plurality of documents on the Internet, such as a blog or a bulletin board, is calculated, and a topic of interest is automatically detected by detecting a word having a high burst degree. Extraction is disclosed.

JP 2009-116440 A

  However, in the above burst detection method, there has been an increase in the number of posts from a relatively narrow range of users who were originally interested in the topic (having the same field of interest), or interested in the topic. It is not possible to determine whether the number of posts from a relatively wide range of users including users who did not have increased. That is, a topic bursted by a wide range of users cannot be detected.

  Therefore, the present invention has been made in view of such a problem, and an object thereof is to provide a burst detection device capable of detecting topics bursted by a wide range of users.

  In order to solve the above-described problem, the burst detection device of the present invention includes posted information history storage means for storing posted information history that is a history of posting time of posted information posted by a user and user identification information for identifying the user. Relevance storage means for storing user relevance that is the relevance of the user to the topic and post information relevance that is the relevance of the post information to the topic; and the post information history stored by the post information history storage means; Based on the user relevance level and post information relevance level stored by the relevance level storage means, post the weighted appearance degree of the topic within the unit time on the time series in the post information history within the unit time for the topic Based on the user relevance of the user who posted the information and the post information relevance of the posted information posted within the unit time for the topic. Comprising a weighted occurrence degree calculating means for calculating you are, burst detection means for detecting a topic that is burst in on the basis of the weighted occurrence of the topic which is calculated by the weighted occurrence degree calculating means, when the series, the.

  If such a configuration is adopted, the weighted appearance degree calculating means determines the weighted appearance degree of the topic within the unit time on the time series in the posted information history of the user who posted the posted information within the unit time for the topic. The weighted appearance degree of the topic calculated by the weighted appearance degree calculating means by the burst detecting means, calculated based on the user relevance degree and the posting information relevance degree of the posted information posted within the unit time for the topic. Based on, a topic bursting in time series is detected. Thus, the weighted appearance level of the topic is calculated based on the user's user relevance level for the topic and the post information relevance level of the post information for the topic. Therefore, for example, the topic weighted appearance degree is calculated to be larger as the user relevance degree is smaller, and the topic weighted appearance degree is larger as the contribution information relevance degree is larger. When there are few users posting about the topic, the weighted appearance of the topic is increased, and as the posted information related to the topic increases, the weighted appearance of the topic is increased, Can be detected. Thereby, it is possible to detect topics bursted by a wide range of users including users with little interest.

  In the burst detection apparatus of the present invention, the weighted appearance degree calculating means may calculate the topic weighted appearance degree as the user relevance degree is smaller. By adopting such a configuration, the topic weighted appearance degree is calculated to be larger as the user relevance degree is smaller. For example, when a user who is less interested in the topic (usually) makes a post related to the topic, the topic Can be detected as a bursted topic. Thereby, it is possible to detect topics bursted by a wide range of users including users with little interest.

  Further, in the burst detection device of the present invention, the burst detection means includes the weighted appearance degree of the topic calculated by the weighted appearance degree calculation means for each unit time included in a predetermined time on the time series. A topic having a weighted appearance level that is equal to or greater than a value based on the calculated average of weighted appearance degrees within a predetermined time may be detected as a burst topic. By adopting such a configuration, it is possible to detect a bursted topic by comparing it with a value based on the calculated average of the weighted appearances within a predetermined time, so that the burst was performed without setting a threshold or the like in advance. Topics can be detected more easily and reliably.

  According to the present invention, topics bursted by a wide range of users can be detected.

It is a functional block diagram of the burst detection apparatus which concerns on embodiment of this invention. It is a hardware block diagram of the burst detection apparatus which concerns on embodiment of this invention. It is a figure which shows the example of a table of contribution information history. It is a figure which shows the example table of topic information. It is a figure which shows the example of a graph of a weighted appearance probability. It is a figure which shows the example of a table of burst information. It is a flowchart which shows the process performed with the burst detection apparatus which concerns on embodiment of this invention. It is a figure which shows the comparison with the technique by the burst detection apparatus concerning embodiment of this invention and the conventional technique.

  Hereinafter, embodiments of a burst detection apparatus according to the present invention will be described in detail with reference to the drawings. In the description of the drawings, the same elements are denoted by the same reference numerals, and redundant description is omitted.

  FIG. 1 is a functional block diagram of a burst detection apparatus 1 according to an embodiment of the present invention. As shown in FIG. 1, the burst detection apparatus 1 includes a posted information history storage unit 10 (posted information history storage unit), an instruction information acquisition unit 11, a topic information calculation unit 12, a topic information storage unit 13 (relevance degree storage unit), A weighted appearance probability calculation unit 14 (weighted appearance degree calculation unit), a burst detection unit 15 (burst detection unit), a burst information storage unit 16 and a burst information output unit 17 are configured.

  FIG. 2 is a hardware configuration diagram of the burst detection device 1. As shown in FIG. 2, the burst detection device 1 shown in FIG. 1 physically includes one or more CPUs 100, a main memory RAM 101 and ROM 102, an input / output device 103 such as a keyboard, mouse and display, The computer system includes a communication module 104 which is a data transmission / reception device such as a network card, and an auxiliary storage device 105 such as a hard disk and a semiconductor memory.

  Each function in FIG. 1 causes the input / output device 103 and the communication module 104 to operate under the control of the CPU 100 by reading predetermined computer software on the hardware such as the CPU 100 and the RAM 101 shown in FIG. This is realized by reading and writing data in the RAM 101 and the auxiliary storage device 105.

  Instead of executing the functions shown in FIG. 1 by a processor such as the CPU 100, all or a part of the functions may be executed by constructing a dedicated integrated circuit. For example, the above functions may be executed by constructing a dedicated integrated circuit for performing image processing and communication control.

  Hereinafter, each functional block of the burst detection apparatus 1 shown in FIG. 1 will be described.

  The posted information history storage unit 10 stores a posted information history that is a time-series history of posted information posted by the user. Posted information is content such as text and images. Specific examples include text (tweets) that expresses the user's situation and notes, emails created by the user, images captured by the user, and users on the Internet. Product purchase data related to products purchased at a shopping site. The posted information history is information in which posted information is associated with time information such as the time and date when the posted information is posted (or created). In the present embodiment, it is assumed that the burst detection device 1 has a microblogging function, and the posted information history storage unit 10 stores tweets posted by a plurality of users.

  FIG. 3 is a diagram illustrating a table example of the posted information history stored by the posted information history storage unit 10. As shown in the table example in FIG. 3, the posted information history includes a “post ID” indicating identification information of each posted information, a “posted content” indicating the content (content) of the posted information, and the user who posted the posted information. “Posting user ID” (user identification information) indicating identification information and “posting date” (posting time) indicating the posting date and time when the posting information was posted by the user are associated with each other.

  In this embodiment, the burst detection device 1 includes the posted information history storage unit 10, but an external server device connected to the burst detection device 1 via the network has the posted information history storage unit 10 (similar function). You may have. In that case, the burst detection device 1 can realize the same processing as the processing described in the present embodiment by transmitting / receiving necessary information and the like to / from the posted information history storage unit 10 via the network. .

  The instruction information acquisition unit 11 acquires instruction information related to an instruction from an administrator or the like of the burst detection device 1. The instruction information includes at least one of information related to a target period (for example, start date / time and end date / time) for calculating topic information described later and information regarding a target period (for example, start date / time and end date / time) to detect a burst described later. But you can. The instruction information acquisition unit 11 may acquire the instruction information from another device via the network and communication module 104, or may acquire it via the input / output device 103 (for example, a keyboard). The instruction information acquisition unit 11 outputs the acquired instruction information to the topic information calculation unit 12 and the weighted appearance probability calculation unit 14.

  First, the topic information calculation unit 12 acquires a posting information history from the posting information history storage unit 10. Specifically, when the instruction information is input from the instruction information acquisition unit 11, the topic information calculation unit 12 extracts and extracts a target period for calculating topic information (described later) included in the input instruction information. The posting information history posted during the target period is acquired from the posting information history storage unit 10. For example, when the target period is indicated by the start date and time and the end date and time, in the post information history table example shown in FIG. 3, a post information history consisting of records having “post date and time” between the start date and end date is extracted. And get. Note that the timing at which the topic information calculation unit 12 acquires the posting information history is not limited to the case where the instruction information is input from the instruction information acquisition unit 11, but may be, for example, regularly (for example, once a week). Good. In this case, the topic information calculation unit 12 uses a preset target period (for example, the past one week) as the target period. It is assumed that the target period set in advance is stored in advance in the burst detection device 1, for example.

  Subsequently, the topic information calculation unit 12 based on the acquired posted information history, the topic of the posted information extracted from each posted information included in the acquired posted information history, and the user's affiliation probability ( The user affiliation probability (user relevance) that is the relevance) and the post information affiliation probability (post information relevance) that is the affiliation probability (relevance) of each post information for each topic are calculated. The topic is information indicating a potential meaning such as a category or field of the posted information, and may be expressed from a word included in the posted information and an appearance probability of the word in the entire target posting. In this embodiment, the calculated topic, user affiliation probability, and post information affiliation probability are collectively referred to as topic information. The topic information calculation unit 12 outputs (stores) the calculated topic information to the topic information storage unit 13 and outputs completion information indicating that the calculation of topic information has been completed to the weighted appearance probability calculation unit 14.

More specifically, the topic information calculation unit 12 calculates a topic by performing topic modeling that is a conventional technique on the obtained posting information history (for a target period for calculating the topic information), User membership probability p (t | u), which is the membership probability that each user u belongs to topic t, and post information membership probability p (t | d), which is the membership probability that each posting information d belongs to topic t. calculate. The affiliation probability that each user u belongs to the topic t indicates the probability that each posting information d posted by each user u is assigned to the topic t. Further, the affiliation probability that each posting information d belongs to the topic t indicates the probability that the topic t is assigned to a word in each posting information d. One example of topic modeling is LDA (Latent Dirichlet Allocation) disclosed in the following non-patent literature. When estimating a topic using LDA, first, a word appearance probability is generated for each topic, and a topic distribution is generated for each post information using the information. Refer to the following non-patent literature for the specific calculation method of the user belonging probability p (t | u) and the posting information belonging probability p (t | d).
Non-patent literature: Blei, David M., Andrew Y. Ng, and Michael I. Jordan. "Latent dirichlet allocation." The Journal of machine Learning research 3 (2003): 993-1022.

  FIG. 4 is a diagram illustrating a table example of topic information calculated by the topic information calculation unit 12. The table example of FIG. 4A shows a list of topics calculated for the posting information history for the target period for calculating the topic information, and includes a “topic ID” indicating topic identification information, and the topic Are associated with “word appearance probability distribution”, which is a word appearance probability distribution. The table example of FIG. 4B shows a table example of information related to the user affiliation probability of the user, “user ID” that is the posting user ID of the user, “topic ID” that indicates the topic identification information, “User affiliation probability” that is the user affiliation probability of the user with respect to the topic is associated. The table example in FIG. 4C shows a table example of information related to the post information affiliation probability of the post information. The “post information ID” which is the post information identification information and the “topic ID” which shows the topic identification information. And “posting information affiliation probability” which is the postal information affiliation probability of the post information d for the topic.

  Returning to FIG. 1, the topic information storage unit 13 stores the topic information input from the topic information calculation unit 12. Note that the topic information storage unit 13 may store only one of the user affiliation probability (information related) and the posting information affiliation probability (information related). The topic information storage unit 13 stores, for example, the topic information shown in the table example of FIG.

  The topic information storage unit 13 may store topic information received from another device via the network instead of the topic information input from the topic information calculation unit 12, or may be input by the administrator of the burst detection device 1. The topic information may be stored. That is, the topic information storage unit 13 may store topic information given (prepared or calculated) in advance. In this case, the function of the topic information calculation unit 12 described above may not be used. That is, for example, the burst detection device 1 may not perform the topic modeling process.

  When the completion information is input from the topic information calculation unit 12, the weighted appearance probability calculation unit 14 is included in the posted information history stored by the posted information history storage unit 10 and the topic information stored by the topic information storage unit 13. Based on the user affiliation probability and post information affiliation probability, the weighted appearance probability (weighted appearance degree) of the topic within the time series unit time in the posted information history is posted within the unit time for the topic. Is calculated based on the user affiliation probability of the user who posted the post information and the post information affiliation probability of the post information posted within the unit time for the topic. In this case, the weighted appearance probability calculation unit 14 may calculate the topic weighted appearance probability to be larger as the user affiliation probability is smaller. This will be specifically described below.

  First, the weighted appearance probability calculation unit 14 acquires topics (lists) included in the topic information stored by the topic information storage unit 13. Here, the weighted appearance probability calculation unit 14 calculates the weighted appearance probability of the topic for each acquired topic. Next, the weighted appearance probability calculation unit 14 extracts a target period for detecting a burst included in the instruction information input from the instruction information acquisition unit 11. Note that a target period for detecting a burst may be one stored in the burst detection device 1 in advance. Next, the weighted appearance probability calculation unit 14 acquires the posted information history posted in the extracted target period from the posted information history storage unit 10. Next, the weighted appearance probability calculation unit 14 acquires a preset unit time l (for example, 5 minutes). It is assumed that the preset unit time l is stored in advance in the burst detection device 1, for example. In addition, the unit time l may be included in the above-described instruction information, and the unit time l may be used.

ここで、Ｄ_ｌは、取得した投稿情報履歴に含まれる投稿情報のうち、抽出した対象期間の単位時間ｌ内の日時に投稿された全投稿情報の集合であり、｜Ｄ_ｌ｜は当該集合に含まれる投稿情報の投稿数である。ｐ（ｔ｜ｕ）は、全投稿情報の集合Ｄ_ｌに含まれる投稿情報ｄを投稿したユーザｕ（取得した投稿情報履歴のうち投稿情報ｄに対応付けられたユーザｕ）がトピックｔに所属するユーザ所属確率ｐ（ｔ｜ｕ）であり、トピック情報格納部１３によって格納されたユーザ所属確率に関する情報のうちユーザｕ及びトピックｔに関するユーザ所属確率を重み付き出現確率算出部１４が抽出することにより得られたものである。ｐ（ｔ｜ｄ）は、全投稿情報の集合Ｄ_ｌに含まれる投稿情報ｄがトピックｔに所属する投稿情報所属確率ｐ（ｔ｜ｄ）であり、トピック情報格納部１３によって格納された投稿情報所属確率に関する情報のうち投稿情報ｄ及びトピックｔに関する投稿情報所属確率を重み付き出現確率算出部１４が抽出することにより得られたものである。 Next, the weighted appearance probability calculation unit 14 calculates the weighted appearance probability wp of each acquired topic t within each unit time l on the time series within the extracted target period by the following equation (1). To do.

Here, D _l is a set of all post information posted on the date and time within the unit time l of the extracted target period among the post information included in the acquired post information history, and | D _l | is the set This is the number of postings of posting information included in. p (t | u) indicates that the user u who posted the posted information d included in the set D ₁ of all posted information (the user u associated with the posted information d in the acquired posted information history) belongs to the topic t. The weighted appearance probability calculation unit 14 extracts the user affiliation probabilities related to the user u and the topic t from the information related to the user affiliation probabilities stored by the topic information storage unit 13. Is obtained. p (t | d) is a post information affiliation probability p (t | d) that the post information d included in the set D _l of all post information belongs to the topic t, and the post stored by the topic information storage unit 13 This is obtained by the weighted appearance probability calculation unit 14 extracting the post information affiliation probabilities related to the post information d and the topic t among the information related to information affiliation probabilities.

Here, “p (t | d) / p (t | u)” included in the numerator in the above formula (1) has a high probability that the posted information d belongs to the topic t, and the posted information d is The lower the probability that the posted user u belongs to the topic t, the larger the weighted appearance probability wp is calculated. Further, “| D _l |” included in the denominator of the above formula (1) means that the weighted appearance probability wp is calculated to be smaller as the number of posts in the unit time l is larger.

FIG. 5 is a diagram illustrating a graph example of the weighted appearance probability wp of a certain topic in the time periods 11 to l15 calculated by the weighted appearance probability calculation unit 14. In the example of the graph shown in FIG. 5, the horizontal axis indicates time series (time), and the vertical axis indicates the value (score) of the weighted appearance probability wp. The wp ₀ and αwp ₀ in the example graph shown in FIG. 5 will be described below.

ここで、Σ_ｌは、取得した抽出した対象期間に含まれる単位時間ｌの合計を示す。 The weighted appearance probability calculation unit 14 is weighted as an average of topic weighted appearance probabilities wp calculated for each unit time l included in a target period (predetermined time) on a time series (detecting a burst). The average appearance probability wp ₀ may be calculated. Specifically, the weighted appearance probability calculation unit 14 calculates the weighted average appearance probability wp ₀ of the topic t by the following equation (2).

Here, Σ _l indicates the total of unit times l included in the acquired extracted target period.

Furthermore, the weighted appearance probability calculation unit 14 acquires a preset threshold value α (for example, 1.5). It is assumed that the preset threshold value α is stored in advance in the burst detection device 1, for example. Then, the weighted appearance probability calculation unit 14 calculates αwp ₀ by multiplying the calculated weighted average appearance probability wp ₀ by the acquired threshold value α.

In the example of the graph shown in FIG. 5, wp ₀ is the weight of the topic calculated for each unit time l included in the target periods l1 to l15 calculated by the weighted appearance probability calculation unit 14 according to the above equation (2). A weighted average appearance probability wp ₀ that is an average of the appearance probabilities wp. Αwp ₀ is a value obtained by multiplying the weighted average appearance probability wp ₀ by α. A method of using wp ₀ and αwp ₀ will be described in the description of the burst detection unit 15 described later.

The weighted appearance probability calculation unit 14 outputs the calculated weighted appearance probabilities wp and αwp ₀ of each topic to the burst detection unit 15. Note that the timing at which the weighted appearance probability calculation unit 14 calculates the weighted average appearance probability is not limited to the case where the completion information is input from the topic information calculation unit 12, for example, (the situation where topic information is calculated in advance) The instruction information may be input from the instruction information acquisition unit 11.

  The burst detection unit 15 detects a topic bursting on the time series based on the weighted appearance probability of the topic calculated by the weighted appearance probability calculation unit 14. The burst detection unit 15 includes a weighted appearance probability that is equal to or higher than a predetermined threshold among the weighted appearance probabilities calculated by the weighted appearance probability calculation unit 14 for each unit time included in the target period on the time series. May be detected as a bursted topic. Here, the predetermined threshold may be a value based on an average of the weighted appearance probabilities of topics calculated by the weighted appearance probability calculation unit 14 for each unit time included in the target period on the time series.

Specifically, the burst detection unit 15 has a weight of αwp ₀ or more input from the weighted appearance probability calculation unit 14 out of the weighted appearance probability wp of each topic t input from the weighted appearance probability calculation unit 14. The topic t with the appearance probability wp is detected as a burst topic. For example, in the graph example illustrated in FIG. 5, the burst detection unit 15 detects four topics with weighted appearance probabilities wp of l8, l9, l10, and l11 that are αwp ₀ or more as burst topics. Here, the detected burst of topic indicates a state in which a user who is not usually interested in the topic is posting on the topic.

  The burst detection unit 15 outputs burst information related to the detected burst topic to the burst information storage unit 16 and the burst information output unit 17. FIG. 6 is a diagram illustrating an example of a burst information table. As shown in the example of the table in FIG. 6, the burst information includes “topic ID” that is identification information of a bursted topic, “weighted appearance probability” that is a weighted appearance probability wp of the topic, and appearance of a word of the topic. The probability distribution “word appearance probability distribution” is associated. Note that the word appearance probability distribution is calculated by performing topic modeling by the topic information calculation unit 12. However, when topic information prepared in advance without using topic modeling is used, this is shown in this table example. Not limited to things.

Note that burst detection of time-series numerical data itself is one of the prior arts. For example, the following non-patent document discloses that a burst period can be detected by defining a cost function or the like in state transition (burst, non-burst).
Non-patent literature: Kleinberg, Jon. "Bursty and hierarchical structure in streams." Data Mining and Knowledge Discovery 7.4 (2003): 373-397.

  The burst information storage unit 16 stores the burst information input from the burst detection unit 15. The burst information storage unit 16 stores, for example, a burst information table example shown in FIG.

  The burst information output unit 17 outputs the burst information input from the burst detection unit 15. More specifically, the burst information output unit 17 transmits the burst information input from the burst detection unit 15 to another device via the network or displays it on the display of the burst detection device 1. The burst information output unit 17 may output the burst information stored by the burst information storage unit 16 when instructed by an administrator of the burst detection apparatus 1 or periodically. Further, when displaying the burst information, the burst information output unit 17 may perform display based on the weighted appearance probability value included in the burst information. For example, the burst information output unit 17 may display information on topics corresponding to the descending order of weighted appearance probability.

  Next, processing of the burst detection method in the burst detection apparatus 1 according to the present embodiment will be described using the flowchart shown in FIG.

  First, instruction information is acquired by the instruction information acquisition unit 11 (step S1, instruction information acquisition step). Next, topic information is calculated by the topic information calculation unit 12 based on the instruction information acquired in S1 and the posting information history stored by the posting information history storage unit 10 (step S2, topic information calculation step). The calculated topic information is stored by the topic information storage unit 13. Next, the weighted appearance probability calculation unit 14 calculates the weighted appearance probability of each topic based on the topic information calculated in S2 or the topic information stored in the topic information storage unit 13 (step S3). , Weighted appearance calculation step). Next, the burst detection unit 15 detects a bursted topic based on the weighted appearance probability calculated in S3 (step S4, burst detection step), and the topic information regarding the detected bursted topic is burst information. Stored by the storage unit 16. Next, the burst information output unit 17 outputs the burst information related to the burst topic detected in S4 or the burst information stored by the burst information storage unit 16 (step S5, burst information output step).

  The processing of the burst detection method in the burst detection device 1 has been described above. When topic information is stored in advance by the topic information storage unit 13, the processes of S1 and S2 are not essential, and the burst detection apparatus 1 may start the process from S3.

  Next, the effect of the burst detection apparatus 1 configured as in the present embodiment will be described.

  According to the burst detection device 1 of the present embodiment, the weighted appearance probability calculation unit 14 determines the weighted appearance probability of a topic within a unit time on the time series in the posted information history within the unit time for the topic. Calculated based on the user affiliation probability of the user who posted the post information and the post information affiliation probability of the post information posted within the unit time for the topic, and by the burst detection unit 15 by the weighted appearance probability calculation unit 14 Based on the calculated weighted appearance probability of the topic, a topic bursting in time series is detected. Thus, the weighted appearance probability of a topic is calculated based on the user's user affiliation probability for the topic and the posting information affiliation probability of the posting information for the topic. Therefore, for example, the topic weighted appearance probability is increased as the user affiliation probability is smaller, and the topic weighted appearance probability is increased as the contribution information affiliation probability is larger. When there are few users posting about the topic, the weighted appearance probability of the topic is increased, and the weighted appearance probability of the topic is increased as the posted information related to the topic is related to the topic. Can be detected. Thereby, it is possible to detect topics bursted by a wide range of users including users with little interest.

  Also, according to the burst detection device 1 of the present embodiment, by calculating the topic weighted probability of occurrence as the user affiliation probability is smaller, for example, users who are less interested in the topic (usually) are concerned with the topic. When the posting is made, it is possible to increase the weighted appearance probability of the topic and detect it as a burst topic. Thereby, it is possible to detect topics bursted by a wide range of users including users with little interest.

  Also, according to the burst detection device 1 of the present embodiment, a burst topic can be detected by comparing the calculated topic with a value based on the average of the weighted appearance probabilities within a predetermined time. It is possible to detect a bursted topic more easily and reliably without setting.

  Although the present embodiment has been described using topics, the present invention is not limited to this. For example, instead of a topic, a keyword, an image, or the like may be used as information indicating the characteristics of the content of posted information.

  As described above, in the burst detection device 1 according to the present embodiment, the past posting information for each user stored in the posting information database 2 is used to classify users into topics. Then, by using the user's affiliation probability in the classified topic, it is determined whether the contributions of users who were not originally interested in are increasing (whether they are showing social spread). That is, burst detection is performed in consideration of the user's affiliation probability in the topic. As a result, it is possible to detect whether a certain topic has become a topic by some users, or whether a topic has become a topic by a wide range of users (topic spreading of users).

FIG. 8 is a diagram showing a comparison between the case where the conventional technique is used and the case where the technique using the burst detection device 1 of the present embodiment is used when detecting a burst of post information regarding rugby. In the conventional method shown in FIG. 8A, the burst degree (score, graph C0) of posting is calculated based on the appearance rate of keywords in the posting information or the appearance rate of topics in all posting information. That is, in the conventional method, only the keyword or topic appearance rate is used as an index, so the attribute (interested field) of the user who posted the posting information is not considered. In the graph C0, (indicating a high incidence topics rugby) the section of B0 that exceeds the threshold value p ₀ is burst.

On the other hand, in the method using the burst detection device 1 of the present embodiment shown in FIG. 8B, the user who posted the posting information does not normally post about a topic, but if it is posting, it is usually Since it is more important than a user who has posted a topic, the burst degree is weighted using the user's topic affiliation probability (field of interest) (considering the degree of interest of the user). Thereby, it can be judged whether the contribution information of the user who was not originally interested is increasing (whether it is showing social spread). A graph C1 shows the above-described weighted graph (overlapped with the graph C0 in order to clarify the difference from the graph C0). In the graph C1, (indicating the user not interested in usual rugby, the high incidence topics rugby) exceeds the threshold value p ₀ (the B0 different) B1 section has burst.

  The burst detection device 1 of the present embodiment can be applied in the fields of search, natural language processing, and burst.

  DESCRIPTION OF SYMBOLS 1 ... Burst detection apparatus, 10 ... Posting information log | history storage part, 11 ... Instruction information acquisition part, 12 ... Topic information calculation part, 13 ... Topic information storage part, 14 ... Weighted appearance probability calculation part, 15 ... Burst detection part, 16 ... Burst information storage unit, 17 ... Burst information output unit.

Claims (3)

A posting information history storage means for storing a posting information history that is a history of posting time of posting information posted by the user and user identification information for identifying the user;
Relevance storage means for storing a user relevance that is a relevance of a user to a topic and a post information relevance that is a relevance of post information to the topic
Based on the post information history stored by the post information history storage means and the user relevance level and the post information relevance level stored by the relevance degree storage means, the topic information in the time series unit time in the post information history A weight for calculating the weighted appearance based on the user relevance level of the user who posted the post information for the topic within the unit time and the post information relevance level of the post information posted for the topic within the unit time Appearance degree calculation means,
Burst detection means for detecting topics bursted in time series based on the weighted appearance degree of topics calculated by the weighted appearance degree calculation means;
A burst detection device comprising:   The burst detection apparatus according to claim 1, wherein the weighted appearance degree calculating unit calculates the topic weighted appearance degree as the user relevance degree is smaller. The burst detection means includes, among the weighted appearance degrees of topics calculated by the weighted appearance degree calculation means for each unit time included in a predetermined time on the time series, within the calculated predetermined time The burst detection device according to claim 1, wherein a topic having a weighted appearance degree that is equal to or greater than a value based on an average of the weighted appearance degrees is detected as a burst topic.

Images