JP6395287B2 - Event detection apparatus and program - Google Patents

Description

  The present invention relates to an event detection apparatus and a program for detecting an event from sequence data.

  Posts on the Internet (such as bulletin board sites and short text posting sites) post on various topics every day. These posts are written in natural language, but many attempts have been made to detect changes in social conditions, incidents and accidents based on such posts.

  In such a situation, one of the required technologies is to detect an abnormal value of the number of writes such as postings via the Internet. One of the conventional methods is a method for detecting a case where the number of writing exceeds a predetermined threshold. Another method is to use a model based on the eigenvalues of the autocorrelation matrix.

  FIG. 13C of Patent Document 1 and paragraph [0116] describe the mutual subspace method. In this method, the eigenvalue problem of the autocorrelation matrix is solved based on the observation data (time series data) that goes back in the past, and the cosine (cos θ) of the angle (θ) between the subspaces is used as the similarity, and the abnormalities of the observation data The value is detected.

JP 2013-041448 A

However, the method for detecting an abnormal value according to the prior art has a problem that sufficient detection accuracy cannot be obtained.
The present invention has been made based on the above problem recognition, and provides an event detection apparatus and program with high accuracy of abnormal value detection.

  [1] In order to solve the above problem, an event detection apparatus according to an aspect of the present invention acquires time series data of the number of utterances, and adds an autoregressive model or an autoregressive moving average model to the acquired time series data. Autoregressive model applying unit that applies and outputs residual time series data obtained as a result of application, and each residual constituting the time series data of residuals output by the autoregressive model applying unit or An event detection unit that detects a location where the absolute value of the residual is larger than a predetermined threshold as an event candidate in the time-series data.

  [2] Further, according to one aspect of the present invention, in the event detection device, the utterance data including the utterance content and time information is acquired, and the utterance content corresponding to a specific topic is acquired from the acquired utterance data. An extraction unit that outputs only the utterance data or only the utterance data that corresponds to a specific topic and corresponds to a specific utterance type, and outputs the pure state utterance data; and the pure state utterance output by the extraction unit A series data generation unit that acquires data and generates time series data of the number of utterances representing the number of the pure state utterance data for each time interval, and the autoregressive model application unit includes the series data Obtain time series data of the number of statements generated by the generation unit, and apply an autoregressive model or an autoregressive moving average model, And wherein the door.

[3] Further, according to an aspect of the present invention, in the event detection device, the event detection unit includes a predetermined number or more consecutive event candidates in the residual time-series data among the detected event candidates. Detecting a location as an event, and detecting as a noise a duration determination unit that detects, as noise, a location where the event candidate continues less than the predetermined number in the time series data of the residual, and detects it as the noise by the duration determination unit A removal processing unit that adjusts the time-series data of the number of utterances so that the absolute value of the residual of the noise portion is reduced.
Note that noise is removed by the removal processing unit. In addition, after the removal processing unit adjusts the time-series data of the number of messages, the autoregressive model application unit performs processing from the point where the autoregressive model or the autoregressive moving average model is applied again.

  [4] Further, according to one aspect of the present invention, in the event detection device, the removal processing unit obtains a degree of contribution to the total amount of noise in the residual time-series data for each location, and the contribution The time-series data of the number of utterances is adjusted in preference to the part where the degree of the message is large.

  [5] Further, according to one aspect of the present invention, as a result of applying a computer to obtain time series data of the number of messages, applying an autoregressive model or an autoregressive moving average model to the obtained time series data, Autoregressive model applying means for outputting the time series data of the residual obtained, each residual constituting the time series data of the residual outputted by the autoregressive model applying means, or an absolute value of the residual is predetermined. Is a program for functioning as event detection means for detecting a location larger than the threshold value as an event candidate in the time-series data.

  According to the present invention, it is possible to detect an event with high accuracy based on a residual of an autocorrelation model or an autocorrelation moving average model. In addition, when only a specific topic or a specific utterance type is extracted and time-series data is generated, the accuracy of model parameter estimation increases, so that an event can be detected with higher accuracy.

It is a block diagram which shows schematic function structure of the event detection apparatus by embodiment of this invention. It is a block diagram which shows the further detailed functional structure of the event detection part by the embodiment. It is the schematic which shows the structure of the comment data by the same embodiment, and a data example. It is the schematic which shows the structure and example of data of the time series data (message number data) by the embodiment. It is a flowchart which shows the operation | movement procedure of the event detection apparatus by the embodiment. It is a flowchart which shows the procedure of the detailed process by the removal process part by the embodiment.

Next, an embodiment of the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram showing a schematic functional configuration of the event detection apparatus according to the present embodiment. In the figure, reference numeral 1 denotes an event detection device. As shown in the figure, the event detection apparatus 1 includes a utterance data acquisition unit 10, an extraction unit 11, a series data generation unit 12, an autoregressive model application unit 13, an event detection unit 14, and an event list output unit 15. It is comprised including.

  The utterance data acquisition unit 10 takes in utterance data from the outside and temporarily stores it in a predetermined format. In order to store the utterance data, for example, a relational database or an XML database (“XML” is an abbreviation of Extensible Markup Language) is used. Here, the utterance data is utterance data posted to a posting service using the Internet. An example of the posting service is a short sentence posting service. In the short text posting service, a user posts a short text from a client terminal device (PC, tablet, smartphone, etc.). There is a predetermined upper limit (for example, about several hundred characters) for the length of a short sentence. The submitted short text is stored in the storage means on the server device side in association with attribute information such as the user account name and the posting date (YYYY / MM / DD hh: mm: ss (year / month / day, hour / minute / second) format). Is done. The server provides the accumulated short text data so that it can be displayed in the user's desired form. At this time, the user can display only posts of a predetermined user account (single or plural) in time series, or can display only posts including a specific keyword in time series. An example of such a short text posting service is Twitter. One post of a short sentence or the like is hereinafter referred to as “speech”.

The extraction unit 11 obtains utterance data including the utterance content and time information, and from the obtained utterance data, only the utterance data having the utterance content corresponding to the specific topic, or the specific utterance corresponding to the specific topic. Outputs pure state utterance data in which only utterance data corresponding to the type is extracted.
The series data generation unit 12 acquires the pure state utterance data output by the extraction unit 11 and generates time series data of the number of utterances representing the number of pure state utterance data for each time interval.

The autoregressive model application unit 13 acquires time series data of the number of messages, applies an autoregressive model or an autoregressive moving average model to the acquired time series data, and obtains a time series of residuals obtained as a result of application Output data.
The event detecting unit 14 determines each residual constituting the time series data of the residual output by the autoregressive model applying unit 13 or a place where the absolute value of the residual is larger than a predetermined threshold as time series data. Detect as event candidate in.

  The event list output unit 15 outputs a list of events detected by the event detection unit 14. Here, the event list is a list of data including at least a time representing a time zone corresponding to the event.

  FIG. 2 is a block diagram showing a more detailed functional configuration of the event detection unit. As illustrated, the event detection unit 14 includes a residual time series data acquisition unit 141, an event candidate time detection unit 142, a duration determination unit 143, a removal processing unit 144, and a detection result output unit 145. Consists of.

The residual time series data acquisition unit 141 acquires residual time series data from the outside (specifically, the autoregressive model application unit 13). This is a residual in the model applied by the autoregressive model application unit 13.
The event candidate time detection unit 142 detects the event candidate time based on the residual time series data acquired by the residual time series data acquisition unit 141. Specifically, event candidate time detection unit 142 detects each time zone (time) as an event candidate when the absolute value of the residual is equal to or greater than a predetermined threshold.

The duration determination unit 143 detects, as an event, a portion of the detected event candidates where the event candidates continue for a predetermined number or more in the time series data of residuals, and the event candidates in the time series data of the residuals are detected. A portion where only less than the predetermined number continues is detected as noise.
The removal processing unit 144 adjusts the time-series data of the number of utterances so that the absolute value of the residual of the noise part becomes small for the part detected as noise by the duration determination unit 143. Adjusting the time-series data of the number of utterances means that the number of utterances is reduced by one in the case of a positive residual, and the number of utterances is increased by one in the case of a negative residual. Adjustment of this data by the removal processing unit 144 has an effect of removing event candidates (that is, noise) that are not true events.

  The detection result output unit 145 supplies a portion detected as an event by the duration determination unit 143 to the outside as a detection result. Specifically, the detection result output unit 145 passes the event detection result to the event list output unit 15.

  FIG. 3 is a schematic diagram illustrating a configuration of the utterance data and a data example. As shown in the figure, the utterance data is tabular data and includes items (digits) of date and time, utterance account, utterance, topic, and utterance type. Each line of the utterance data corresponds to one utterance. The date and time is the date and time when the utterance is posted, and is data expressed in year, month, day, hour, minute and second. The comment account is the account name of the user who posted the comment. The utterance is text data of a posted sentence. In other words, the message column stores the data of the message content. The topic is data of a label representing the topic to which the utterance belongs. The message type is label data representing the type of the message.

  Here, the message type will be described. The utterance type is a classification within a specific topic and is based on the content of the utterance. For example, an impression regarding a certain event is classified into three types of utterances of “affirmation”, “denial”, and “request” (this is referred to as an impression type for convenience). Opinions on a certain policy are classified into two types of remarks, “support” and “non-support” (this is called a support type for convenience). In a posting service that posts and browses in real time via the Internet, the same type of opinion tends to spread within the user's community.

  In the figure, as a data example, records for a total of six lines are shown. Data on the first line (date and time is “February 8, 2014 8:12:44”) and data on the second line (date and time is “February 8, 2014 8:12:47”) The topic is “XX tax rate increase”. The message type on the first line is “No”, and the message type on the second line is “Yes”. Records shown in the third and subsequent lines correspond to statements about other topics.

At the stage where the utterance data is acquired from the outside (for example, the data distribution server device of the short posting site), the data is included only in the date / time, utterance account, and utterance columns, and the topic and utterance type columns are blank. . After the topic classification, data is written in the topic column. Further, after the classification of the message type is performed, the data is written in the column of the message type. In the state shown in the figure, topics are mixed.
After data is written in the topic column, only data on a specific topic can be selected and extracted by data search using conditions. In addition, after data is written in the message type column, only data of a specific message type can be selected and extracted by data search using conditions. In addition, after data is written in both the topic field and the message type field, only data of a specific topic and a specific message type can be selected and extracted by data search using conditions. it can. Data extracted only for a specific topic or type of speech can be used as pure state speech data.

  FIG. 4 is a schematic diagram showing a configuration of time-series data (message count data) and a data example. As shown in the figure, this time-series data represents the number of utterances of a specific topic and a specific utterance type for each time zone (time interval). This time-series data has a tabular structure and includes items such as time zone (date and time), topic, statement type, and number of cases. Each row of the time-series data shown in the figure corresponds to a combination of time zone, topic, and message type.

The time zone (date and time) is date and time data representing a time zone having a predetermined length (width). The width of the time zone is appropriately determined according to the nature of the data to be analyzed, the nature of the topic of speech, and the like. In normal cases, the width of the time zone is about 1 minute to about 1 day. In the example shown in the figure, the width of the time zone is 5 minutes, and the start time of each time zone is represented as the representative date and time of the time zone.
The topic is a label representing the topic classified and extracted by the extraction unit 11.
The utterance type is a type obtained by classifying the utterance contents as described above. For example, there are two types of speech in a specific topic: “affirmation” and “denial”.
The number of cases is numerical data representing the number of pieces of speech data corresponding to the time zone (date and time), the topic, and the speech type.

In the illustrated data example, the width of the time zone is 5 minutes. The first and second lines of the data example are data related to the time zone from “8:10 on February 8, 2014” to “8:15” on the same day. As representative, data “February 8, 2014 8:10” is stored in the time zone (date and time) column. The third and fourth lines are data relating to the time zone from “8:15 on February 8, 2014” to “8:20” on the same day. In the time zone (date and time) column, data “February 8, 2014, 8:15” is stored. In all the data of this data, the topic is “XX tax rate increase”. That is, this time series data represents the time series of the number of utterances related to the tax rate increase of “XX tax” extracted from the utterance data. The message types in the first and third lines are “No”, and the message types in the second and third lines are “Yes”. Further, the number of utterances for each utterance type in each time zone is stored in the number of cases column.
The time series data is generated by the series data generation unit 12.

Next, the processing procedure will be described.
FIG. 5 is a flowchart showing an operation procedure of the event detection apparatus. Hereinafter, the operation of the event detection apparatus 1 will be described with reference to this flowchart.
First, in step S101, the speech data acquisition unit 10 acquires speech data from the outside. This message data is shown in FIG.

  Next, in step S102, the extraction unit 11 extracts only the utterances of the specific topic from the utterance data. Further, the extraction unit 11 classifies the extracted topic message data for each message type. The extraction unit 11 writes data to the topic item and the comment type item in the comment data based on the classification and extraction results. As a result of such classification, data in a state where only a specific topic and a specific message type are extracted is referred to as pure state message data.

In order to extract only the utterances of a specific topic, an automatic classification technique using a natural language processing technique can be used. This technology itself is an existing technology as described in the literature [Michael W. Berry, Survey of Text Mining, 2004 edition, Springer]. It is extracted by keyword setting and rules (combination of multiple conditions, etc.) ) In combination with extraction. In addition, in order to extract only utterances related to a specific broadcast program such as a television, the above automatic classification technique is used using keywords related to the broadcast program.
In addition, in order to extract only a specific utterance type, as in the literature [Michael W. Berry, Survey of Text Mining, 2004 edition, Springer], it is possible to use the reputation analysis technology that is an existing technology. The reputation analysis technique classifies, for example, a multidimensional vector representing the frequency of words appearing in a statement in the vector space. In addition, the reputation analysis technique learns how to classify in the vector space by a machine learning technique using teacher data as necessary.

  In step S103, the series data generation unit 12 generates time series data based on the classified message data. An example of the time series data generated by the series data generation unit 12 is as shown in FIG. Specifically, the series data generation unit 12 generates the time series data by counting the number of message data extracted by the extraction unit 11 for each time zone, each topic, and each message type.

Next, in step S <b> 104, the autoregressive model application unit 13 applies the autoregressive model to the time series data generated by the series data generation unit 12. As a result, the autoregressive model application unit 13 calculates data of time series (ε ₁ , ε ₂ ,...) Of residual values. Here, when the autoregressive model application unit 13 applies an AR model (autoregressive model, autoregressive model) and when an ARMA model (autoregressive moving average model) is applied, there are two ways An example will be described.

(1) When AR model is applied An autoregressive model of order p (p is a natural number) is expressed by the following equation (1).

In the equation (1), φ ₁ ,..., Φ _p are model parameters. C is a constant term. ε _t is a residual (error) term. That is, in the model shown in equation (1), the value _{X t} of the time series data at discrete time t, p number of values _X t-1 immediately before the same series, _..., weighted for _{X t-p} The sum is a constant term and a residual term.

The autoregressive model application unit 13 analyzes the time series data supplied from the series data generation unit 12 according to the model shown in Expression (1). The process itself of analyzing the series by applying the autoregressive model is based on the existing technology. Specifically, the autoregressive model application unit 13 obtains a solution of simultaneous multiple equations based on the supplied time series data or performs maximum likelihood estimation of the solution, and as a result, a sequence of residual terms ε ₁ , epsilon _2, obtained and ..., and the constant term c, parameter phi _1, ..., the value of phi _p.

(2) When applying the ARMA model A model combining the autoregression of the order p and the moving average of the order q is expressed by the following formula (2) (p and q are natural numbers).

In equation (2), φ ₁ ,..., Φ _p and θ ₁ ,..., Θ _q are model parameters. ε _t is a residual (error) term. That is, in the model shown in equation (2), the value _{X t} of the time series data at discrete time t, p number of values _X t-1 immediately before the same series, _..., weighted for _{X t-p} The sum, the weighted sum of q immediately preceding residual terms ε _t−1 ,..., Ε _{t-q in} the same series, and the value of the residual term at the time t are added.

The autoregressive model application unit 13 analyzes the time series data supplied from the series data generation unit 12 according to the model shown in Expression (2). The process itself of analyzing the series by applying the autoregressive moving average model is based on the existing technology. Specifically, the autoregressive model application unit 13 obtains a solution of simultaneous multiple equations based on the supplied time series data or performs maximum likelihood estimation of the solution, and as a result, a sequence of residual terms ε ₁ , ε ₂ ,..., parameters φ ₁ ,..., φ _p , and θ _{1 ,.}

As described above, the autoregressive model application unit 13 calculates time series data (ε ₁ , ε ₂ ,...) Of residual values. In both the AR model and the ARMA model, the residual ε _t (t = 1, 2,...) Follows a normal distribution with an average value of 0. In other words, by appropriately setting the threshold value e _th, the absolute value of ε _{t |} ε _t | it is the smaller can exceed the threshold e _th. That is, a time zone in which | ε _t | exceeds the threshold value e _th can be regarded as a candidate event occurrence time zone.

In step S105, the event detection unit 14 determines that the absolute value | ε _t | of the residual exceeds the threshold e _th based on the time series data of the residual value supplied from the autoregressive model application unit 13. All time zones are detected. In other words, the event detection unit 14 obtains all t that satisfy | ε _t |> e _th . Such a time zone is referred to as an event candidate for convenience.
More specifically, the residual time series data acquisition unit 141 takes in time series data of residual values from the autoregressive model application unit 13. Then, the event candidate time detection unit 142 detects all event candidate time zones.

Next, in step S106, the event detection unit 14 (specifically, the duration determination unit 143) determines whether there is an event candidate detected in step S105 that has not exceeded the threshold for a predetermined time. Determine whether. In the present embodiment, the event detection unit 14 makes this determination based on whether there is a sequence whose number of consecutive sequences exceeding the threshold value is less than a predetermined value.
For example, let us consider a case where the event candidate is recognized as an event when the time zone width is 5 minutes and the event candidate (exceeding the threshold) for 20 minutes or more continues. At this time, if four or more time-series event candidates continue, it is recognized as an event. If only three or less time-series event candidates continue, the event candidate is not an event, but “ It is recognized that it is “a part that does not continue for a predetermined time”.
If there is a portion where exceeding the threshold does not continue for a predetermined time (step S106: YES), the process proceeds to step S107. On the other hand, when there is no portion where the threshold value is not exceeded for a predetermined time (step S106: NO), the process proceeds to step S108.

  Next, when the process proceeds to step S107, in the same step, the event detection unit 14 (specifically, the removal processing unit 144) performs data processing for removing a residual corresponding to an event candidate having a short duration. I do. This process acts in a direction that the number of event candidates with a short duration decreases. In other words, by this process, the event detection unit 14 performs data removal (updating) in a direction in which the degree of exceeding the threshold value becomes smaller at a portion where the exceeding of the threshold value does not continue for a predetermined time. Details of the processing in this step will be described later with reference to another flowchart (FIG. 6). After finishing the process of this step, it progresses to the process of step S104 again. As a result, the autoregressive model is again applied to the time-series data after being removed (updated) in this step.

  When the process proceeds from step S106 to step S108, in the same step, the event detection unit 14 (specifically, the detection result output unit 145) outputs the time information of the detected event to the event list output unit 15. To do. The precondition that the control shifts to this step is that, in step S106, it is determined that there is no portion where exceeding the threshold does not continue for a predetermined time. In other words, at the point of time when control is transferred to this step, all points where the threshold value is exceeded (time zones) are all parts that are recognized as events. In other words, the event detection unit 14 provides a list of pairs of the time of the place where the threshold is exceeded (time representative of the time zone) and the time-series data associated with the time to the event list output unit 15. Pass. Then, the event list output unit 15 outputs the data of those times and the time-series data (count of utterances) associated with those times to the outside. Then, when the process of this step is completed, the event detection apparatus 1 ends the process of the entire flowchart.

  FIG. 6 is a flowchart showing a detailed processing procedure by the removal processing unit 144. The process shown in the figure is the process of step S107 in the process procedure shown in FIG. Hereinafter, the process for removing an event having a short duration will be described with reference to this flowchart.

First, in step S201, the removal processing unit 144 in the event detection unit 14 acquires data of a list of event candidates with a short duration. This data is intended to be supplied by the continuation time determination unit 143, as already mentioned, a time period during which the absolute value of the residual exceeds the threshold value e _th, and continued for more than such a threshold It is a list of time zones in which the running time is less than a predetermined time. Specifically, in this step, the removal processing unit 144 acquires a list of pairs of the time representing the corresponding time zone and the value of the number of utterances in the time zone.

Next, in step S202, the removal processing unit 144 obtains a contribution degree for each piece of utterance data for each event candidate included in the acquired data. Here, the degree of contribution to the sum of the square errors of the model represented by Expression (1) or Expression (2) and the actual measurement value (the number of tweets) is referred to as contribution. In other words, for each time zone that has a residual to be removed (a short time residual that is not recognized as an event), the contribution is the time series data as a whole when the number of statements in that time zone is reduced by one. This is the degree to which the total amount of square error is reduced. In order to obtain the degree of contribution to the residual of each time zone targeted by the removal processing unit 144, after actually reducing the number of utterances as described above, the autoregressive model applying unit 13 again receives the autoregressive model. Is calculated, how much the total absolute value of residuals to be removed is reduced. Then, the utterance data having the largest contribution is selected.
In the above-described processing, the sum of squared errors is the sum of all target time zones (time zones that exceed the threshold but are not events).

  Next, in step S203, the removal processing unit 144 reduces the utterance data having the largest contribution (as selected in step S202) by one. This processing is performed in order to remove an utterance having a large degree of contribution to an amount regarded as an event (total sum of square errors) in an utterance that is not recognized as an event.

  As described above with reference to FIGS. 5 and 6, the event detection apparatus 1 according to the present embodiment classifies topics and utterance types in advance, and applies the AR model or the ARMA model as pure utterance data. To do. Further, the event detection apparatus 1 recognizes, as an event candidate, a time zone in which the residual (absolute value) when the AR model or the ARMA model is applied exceeds a predetermined threshold. Further, the event detection apparatus 1 determines whether it is a true event to be detected or noise to be removed depending on whether the duration time of the authorized event candidate is long or short. Moreover, the event detection apparatus 1 removes the noise for the time zone of the event candidate recognized as the noise to be removed.

  In this embodiment, topics and utterance types are classified in advance, and only a specific topic and utterance type data (pure state utterance data) is extracted, and then a plurality of models are generated by generating time-series utterance number data. It is possible to obtain time-series data of the state of a single model, not time-series data of a state of complex. Since the AR model or the ARMA model is applied to such time-series data based on a single model, that is, time-series data having a good property, modeling errors can be reduced. That is, it is possible to detect an event using time series data close to an ideal state assumed by the model as a processing target. If two topics are progressing simultaneously and a single model is applied without separating the topics, an error will occur.

  Further, in the present embodiment, similar to the topic classification, the utterance type (positive or negative for a certain event or the like, whether it is supported or not supported, etc.) is classified and the utterance is classified. Since the time-series data of the number of utterances is generated after the data is separated by the utterance type, it is possible to perform more accurate analysis and more accurate event detection in the same manner as the above-described topic classification.

  In the prior art, the AR model or the ARMA model has not been applied because of the large modeling error. In the present embodiment, as described above, since only the data of the specific topic and the message type is extracted and then the data of the number of messages in the time series is generated, it is possible to satisfactorily apply the AR model or the ARMA model. Become.

  Further, in this embodiment, since the AR model or the ARMA model is applied, the residual (when the AR model or the ARMA model is applied) is not simply recognized as an event when the timing exceeds a predetermined threshold. Timing (time period) when the absolute value) exceeds a predetermined threshold is recognized as an event candidate. This makes it possible to perform more accurate event detection.

  Further, in the present embodiment, it should be regarded as a true event or noise depending on whether the duration of the time zone (event candidate) whose residual (absolute value) exceeds a predetermined threshold is long or short. Determine whether it is an abnormal value. In other words, a time zone in which the residual is a burst, that is, a time zone of an event candidate with a long duration is recognized as an event. Conversely, a residual that is not a burst, that is, a time zone of a short event candidate is recognized as noise. Thereby, it is possible to prevent temporary abnormal noise from being detected as an event.

  In the present embodiment, the noise is removed from the time zone of the event candidate recognized as noise. Specifically, data processing equivalent to the case where the utterance is removed is performed in descending order of contribution to noise, and such removal is repeated until there is no noise (event candidate with a short duration). As a result, a result that closely matches the model can be obtained.

  As described above, in this embodiment, since the estimation accuracy of the parameters of the time series model is improved, the residual of the part that can be explained by the model (the difference between the model and the actual time series data) is reduced. It is expected to obtain the effect. If the amount by which the residual decreases due to this improvement in estimation accuracy is constant regardless of the time zone, the residual peak will appear more clearly. Also, even if the amount by which the residual decreases due to this improvement in estimation accuracy is not strictly constant, if the residual other than multiplicative noise decreases, the residual peak (multiplicity noise peak) It will appear more clearly. As a result, there is an effect that the detection accuracy of a portion that cannot be explained by autocorrelation is improved. Thereby, the event about a speech is detected accurately. In other words, the event detection device 1 can be used to detect changes in social situations, incidents / accidents, etc. corresponding to such events, and the detection accuracy can be improved.

  Note that the function of the event detection apparatus in the above-described embodiment may be realized by a computer. In that case, a program for realizing this function may be recorded on a computer-readable recording medium, and the program recorded on this recording medium may be read into a computer system and executed. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer-readable recording medium” refers to a storage device such as a flexible disk, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, a “computer-readable recording medium” dynamically holds a program for a short time, like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory inside a computer system serving as a server or a client in that case may be included, and a program that holds a program for a certain period of time. The program may be a program for realizing a part of the above-described functions, or may be a program that can realize the above-described functions in combination with a program already recorded in a computer system.

  Although the embodiment has been described above, the present invention can also be implemented in the following modified example.

(Modification 1) For example, in the above embodiment, when the residual is both positive and negative, it is detected as an event candidate when the absolute value of the residual exceeds a threshold (| ε _t |> e _th ). However, this may be changed so that the event candidate is detected only when the residual is positive and the residual exceeds a threshold (ε _t > e _th ). When this modification is implemented, the event detection apparatus 1 detects only events in a time zone in which the residual is positive, and does not detect events in a time zone in which the residual is negative. Become. This modification is preferable when such a detection method is desired depending on the purpose.

(Modification 2) In the above embodiment, the event detection device 1 has the configuration shown in FIG. In the second modification, the event detection device does not include the utterance data acquisition unit 10, the extraction unit 11, and the sequence data generation unit 12 in the configuration illustrated in FIG. The event detection apparatus includes an autoregressive model application unit 13, an event detection unit 14, and an event list output unit 15. In this case, series data is generated by another external device, and the generated series data is read by the autoregressive model application unit 13 and processed.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and different designs may be used without departing from the gist of the present invention.

  The present invention can be used for social analysis and the like. The present invention can be used to detect, for example, changes in social situations, sudden accidents / incidents, and rapid spread of specific information via the Internet. In the broadcasting business, it can be used for program production planning or as an opportunity for news reporting. For public institutions such as the government, it can be used for social policy planning.

DESCRIPTION OF SYMBOLS 1 Event detection apparatus 10 Statement data acquisition part 11 Extraction part 12 Sequence data generation part 13 Autoregressive model application part 14 Event detection part 15 Event list output part 141 Residual time series data acquisition part 142 Event candidate time detection part 143 Duration determination Unit 144 removal processing unit 145 detection result output unit

Claims (4)

Autoregressive model application that obtains time series data of the number of remarks, applies autoregressive model or autoregressive moving average model to the acquired time series data, and outputs residual time series data obtained as a result of application And
Detect each residual constituting the time series data of the residual output by the autoregressive model application unit or a place where the absolute value of the residual is larger than a predetermined threshold as an event candidate in the time series data An event detector to perform,
A event detection device you equipped with,
The event detection unit
Among the detected event candidates, a portion where the predetermined number of event candidates continue in the time series data of the residual is detected as an event, and the event candidates are less than the predetermined number in the time series data of the residual A duration determination unit for detecting non-continuous portions as noise,
For the portion detected as the noise by the duration determination unit, a removal processing unit that adjusts the time-series data of the number of utterances so that the absolute value of the residual of the noise portion is small,
Event detection device characterized by comprising a. The removal processing unit obtains the degree of contribution to the total amount of noise in the residual time-series data for each location, and gives priority to the location where the degree of contribution is large, the time-series data of the number of messages. adjust,
The event detection apparatus according to claim 1 , wherein: Remarks data including remark contents and time information is acquired, and from the acquired remark data, only the remark data having the remark contents corresponding to a specific topic or corresponding to a specific topic and corresponding to a specific remark type An extractor that extracts only said utterance data and outputs pure state utterance data;
Obtaining the pure state utterance data output by the extraction unit, and a sequence data generation unit for generating time series data of the number of utterances representing the number of the pure state utterance data for each time interval;
Further comprising
The autoregressive model application unit acquires the time series data of the number of messages generated by the series data generation unit, and applies an autoregressive model or an autoregressive moving average model.
The event detection apparatus according to claim 1 or 2 , wherein Computer
The event detection device according to any one of claims 1 to 3 ,
Program to function as.

Images