JP2011053776A - System, method and program for generating story - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To generate a story text showing a causal relation for a useful environmental change among actually-occurred environmental changes by a minimum user burden. <P>SOLUTION: An extraordinary event determination unit 13 selects an extraordinary event based on coincidence between features of a time-series change related to situation occurrence data of a newly detected event and features of a time-series change related to situation occurrence data of past events detected in the same past time window which is acquired from an event DB 12. A query generation unit 21 generates query data including these past events as answer candidates from event information for past events correlated with the selected extraordinary event, and a query processing unit 23 identifies a cause event that is the source of the extraordinary event from answers to the query data from users, and generates explanatory texts of the cause event and the extraordinary event. A story generation unit 26 generates a story text from the explanatory texts. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to an environment change detection technique, and more particularly to a technique for generating text information indicating an environment change that has actually occurred based on a detection event detected by a sensor.

  In recent years, research has been progressing to realize environmental intelligence that supports people by using ubiquitous sensor networks to detect environmental changes that occur in the real environment as events and presenting the event contents to people. Yes. In such environmental intelligence, a technique for providing a detected event to a person that is easier to understand, converting the content of the detected event into an electronic text, and posting it on a blog has been proposed (for example, non-patent literature). 1 etc.). However, this technique does not consider the causal relationship between events, and it may be difficult to easily understand the event contents from the generated electronic text.

  On the other hand, as a technique for converting the contents of various events that occur in the real environment into electronic text based on the causal relationship between the events, that is, as a technique for converting the contents of the event into a story, a technique using images has been proposed (for example, non- (See Patent Document 2). This technology makes it possible to record past operations, share experiences, and experience experiences by recording all the operations of the user who uses the PC or combining the shooting and recording of the digital camera. We are also developing an application for manually creating a slideshow-style story using recorded moving images.

In addition, a technique for mounting a device that can easily create a story based on a photograph taken has been proposed (see, for example, Non-Patent Document 3). This device has a liquid crystal display and a microphone, and can add a story read aloud to a picture displayed on the liquid crystal display.
In addition, a system has also been proposed in which a record relating to the user's behavior of the day (such as who met or where he / she went) is recorded in a log composed of electronic text and presented to the user as a story using icons (for example, non- (See Patent Document 4). In this system, an event that a user meets another user is represented by an icon representing each user.

Takuya Maekawa, Yutaka Yanagisawa, Takeshi Okadome, Towards environment generated media: object-participation-type weblog in home sensor network, Proceedings of the 16th international conference on World Wide Web, pp.1267--1268, 2007. Jim Gemmell, Gordon Bell and Roger Lueder, My Life Bits: a personal database for everything, Communications of the ACM, pp. 88-95, Vol. 49, Issue 1, 2006. Marko Balabanovic, Lonny L. Chu, and Gregory J. Wolff, Storytelling with digital photographs, Proc. CHI 2000, pp. 564-571, 2000. M. Lamming and M. Flynn, "Forget-me-not: Intimate computing in support of human memory," Proc. FRIEND21: International Symposium on Next Generation Human Interface, 1994, pp. 125-128. Riessman, Narrative Analysis, saga Publications, 1993. E.M.Tapia, S.S.Intille, and K. Larson, Activity recognition in the home using simple and ubiquitous sensors, Proc. Pervasive 2004, pp. 158-75, 2004. J. Sohn, N.S. Kim, and W. Sung, A statistical model-based voice activity detection, IEEE Signal Processing Letters, 6, pp. 1-3, 1999.

However, in such a conventional technique, when a causal relationship of an event detected as an environmental change is made into a story, a story sentence is created manually or semi-automatically based on a photographed image, photograph, or log. In order to create a story sentence, there has been a problem that humans must be actively involved, which increases the burden on the user.
In daily life, the number of events detected by ubiquitous sensors installed in the environment is extremely large, and the contents vary from trivial to serious. For this reason, even if the space for detecting events is limited, there is a problem in that useless information increases when story sentences are created for all events.

  The present invention is to solve such a problem, and a story generation technique capable of generating a story sentence showing a causal relationship regarding a useful environmental change among environmental changes that have actually occurred with a small user burden. It is intended to provide.

  In order to achieve such an object, a story generation system according to the present invention is a story generation system that generates a story sentence that indicates a causal relationship of an environmental change that has actually occurred based on a detection event detected by a sensor. An event detection unit that accumulates occurrence status data relating to detection events sequentially detected by a sensor in time series together with explanatory information indicating the contents of the detection event as event information for each time window of a predetermined time length, and an event The coincidence between the characteristics of the time series change regarding the occurrence status data of the detected event newly detected by the detector and the characteristics of the time series change about the occurrence status data of the past event detected from the event database in the past same time window The detected event is based on the past event Is determined by the extraordinary event determination unit that accumulates event information of detected events determined as extraordinary events in the extraordinary event database and the extraordinary event determination unit To select past events correlated with unusual events from the event database and unusual event database, and to inquire about the cause of the unusual events based on the event information of these past events, with these past events as answer candidates The cause generator that generates the question data, and the cause event that causes the occurrence of the extraordinary event based on the answer from the user to the inquiry based on the question data generated by the question generator Explanatory text indicating the cause of the environmental change from the event information of the cause event A story sentence that shows the causal relationship of environmental changes by combining a question processing unit that generates an explanatory text showing the result of an environmental change from event information of an extraordinary event and an explanatory text generated by the question processing unit And a story generation unit for generating

  At this time, a detection vector having an occurrence state related to detection events sequentially detected in time series in an arbitrary time window by an extraordinary event selection unit and a plurality of past vectors related to past events obtained in the past in the same time window The principal component analysis processing is performed on the above, and the dimension loss processing of the detection vector is performed based on the obtained main principal component to calculate the information loss rate for the detection vector, and the information loss rate and the loss rate threshold are calculated. You may make it determine whether the said detection vector is an unusual event by comparing with a value.

  Furthermore, when the information loss rate is less than or equal to the loss rate threshold value in the extraordinary event selection unit, the detection vector and the past vector are projected onto the main principal component axis, and the projection is performed for each of these principal component axes. Generate a mixed normal distribution of values, and compare the probability density related to the projected value of the detected vector in the mixed normal distribution with the probability density threshold for each of these principal principal component axes, thereby making the detected vector an unusual event. It may be determined whether or not.

  The story generation method according to the present invention is a story generation method used in a story generation system for generating a story sentence indicating a causal relationship of an environmental change that has actually occurred based on a detection event detected by a sensor. An event detection step for accumulating generation status data relating to detection events sequentially detected by the sensor in time series by the sensor together with explanatory information indicating the contents of the detection events in the event database as event information for each predetermined time window And the non-daily event determination step, the characteristics of the time series change regarding the occurrence status data of the detected event newly detected in the event detection step, and the occurrence status data of the past event detected in the past same time window acquired from the event database Time series Based on the coincidence with the characteristics of the event, it is determined whether or not the detected event is an extraordinary event deviating from the past event, and event information of the detected event determined as an extraordinary event is accumulated in the extraordinary event database. The extraordinary event determination step and the question processing unit select past events correlated with the extraordinary event determined in the extraordinary event determination step from the event database and the extraordinary event database, and based on the event information of these past events The question generation step for generating the question data for inquiring about the cause of the occurrence of the extraordinary event using these past events as answer candidates, and the question processing unit to the user based on the question data generated in the question generation step Based on user responses to Identifies the cause event that causes the occurrence of the daily event, generates an explanatory text that indicates the cause of the environmental change from the event information of the cause event, and generates an explanatory text that indicates the result of the environmental change from the event information of the unusual event And a story generation step in which the story generation unit generates a story sentence indicating the causal relationship of the environmental change by combining the explanation sentences generated in the question processing step.

  At this time, in the extraordinary event selection step, a detection vector having an occurrence state related to detection events sequentially detected in a time series in an arbitrary time window, and a plurality of past vectors related to past events obtained in the past in the same time window The principal component analysis processing is performed on the above, and the dimension loss processing of the detection vector is performed based on the obtained main principal component to calculate the information loss rate for the detection vector, and the information loss rate and the loss rate threshold are calculated. You may make it determine whether the said detection vector is an unusual event by comparing with a value.

  Furthermore, if the information loss rate is less than or equal to the loss rate threshold value in the extraordinary event selection step, the detection vector and the past vector are projected onto the main principal component axis, and the projection is performed for each main principal component axis. Generate a mixed normal distribution of values, and compare the probability density related to the projected value of the detected vector in the mixed normal distribution with the probability density threshold for each of these principal principal component axes, thereby making the detected vector an unusual event. It may be determined whether or not.

  Further, the program according to the present invention is a program for causing a computer to function as each part constituting any one story generation system described above.

According to the present invention, the causal relationship of the environmental change can be achieved by a very simple user operation of selecting a cause event for a result event including a newly detected detection event from among answer candidates presented as question data. A story sentence indicating can be generated. In addition, the user automatically selects a non-daily event that does not occur in daily life from a vast number of events that occur in daily life, and automatically generates a story sentence that uses this non-daily event as a result event. Can be generated.
Therefore, according to the present embodiment, it is possible to generate a story sentence indicating a causal relationship regarding a useful environmental change among environmental changes that have actually occurred with a small user burden.

It is a block diagram which shows the structure of the story production | generation system concerning this Embodiment. It is explanatory drawing which shows the relationship between an event and a story. It is a flowchart which shows the story production | generation operation | movement of the story production | generation system concerning this Embodiment. It is a flowchart which shows an information loss rate calculation process. It is explanatory drawing which shows a part of information loss rate calculation process. It is a flowchart which shows a probability density calculation process. It is a flowchart which shows an extraordinary event determination process. It is a flowchart which shows a question production | generation process. It is explanatory drawing which shows the calculation method of the certainty factor used with an a priori algorithm. It is a flowchart which shows a question process. It is an example of a presentation screen of question data. It is an example of generation of a story sentence.

Next, an embodiment of the present invention will be described with reference to the drawings.
[Story generation system]
First, a story generation system according to the present embodiment will be described with reference to FIG. FIG. 1 is a block diagram showing a configuration of a story generation system according to the present embodiment.

  The story generation system 10 has a function of generating a story sentence indicating a causal relationship of an actual environmental change based on detection events detected by various sensors constituting the sensor network 30.

The sensor network 30 is connected to the story generation system 10 via a communication line 31 and is configured by a so-called ubiquitous sensor that detects the movement of a person or an object with a physical quantity or an image. The detection result of the sensor is used as an event. It has a function of outputting to the story generation system 10.
The Weblog system is connected to the story generation system 10 via the communication network 40, and posts the story sentence to the blog in response to a posting instruction including the story sentence transmitted from the story generation system 10, thereby enabling the communication network 40 has a function of publishing. In addition, the story sentence transmitted from the story generation system 10 can also be posted on a general Web page, a bulletin board, etc. in addition to posting to a blog. In addition, the story sentence transmitted from the story generation system 10 may be output as a printed matter.

[Basic concept of the present invention]
The root of the story is to set up multiple events. According to Riesman (see Non-Patent Document 5 and the like), there are three types of methods for creating a plurality of events: an chronological arrangement, a consequent arrangement, and an episodic arrangement. Arrangement by chronological order means staging the events in the order in which they occurred, and consequent arrangement means staging the events that occurred under a causal relationship. An episodic arrangement is like a combination of chronological and consequential arrangements, but the key is to associate events with the same subject.

  Of these three types of bracing methods, the present invention focuses on the consequent arrangements that are causal. The reason for this is that if events that are elements of different stories are mixed in the chronological order, it will not be realized as a story simply by arranging the events in a time sequence. On the other hand, in the case of an episodic arrangement, it is difficult to grasp what the subject of each event is.

  FIG. 2 is an explanatory diagram showing the relationship between events and stories. When a human writes a diary, it is thought that the event to be recorded is first conceived and then the reason why the event has occurred is explored. In the present invention, a story is created by connecting a cause event that is a cause to a detected event, that is, a result event.

Moreover, in our daily life, a huge number of events occur, and the contents vary from trivial to serious. For this reason, even if the space to be detected is limited, if a story is created and presented for all events that occur in the space, it is considered that useless information increases.
Therefore, the inventors conducted a preliminary questionnaire to the subjects in order to know what kind of event should create and present a story.

In this pre-questionnaire, questions were asked such as "What do you want me to record" and "How convenient is it?" As a result of the preliminary questionnaire, it was found that there were many subjects who wanted to find something different from usual by looking at the recorded data, such as “too much medicine” and “forgetting to clean the aquarium”.
In the present invention, based on such a questionnaire result, only non-daily events that do not occur in daily life are selected as result events, and the causal relationship of the actual environmental change is pursued from the result events. The story sentence which shows the said causal relationship is produced | generated.

  For this reason, the present embodiment is characterized by the characteristics of the time series change regarding the occurrence status data of the newly detected detection event and the time related to the occurrence status data regarding the past event sequentially detected in time series in the same time window in the past. Based on the coincidence with the characteristics of the series change, it is determined whether or not the detected event is an unusual event deviating from the past event, and past events correlated with the determined unusual event are selected. Based on the event information of the event, generate the question data for inquiring about the cause of the occurrence of the extraordinary event with these past events as answer candidates, and based on the user's response to the inquiry to the user based on the question data Identify the cause event that caused the extraordinary event, and Generate explanatory text showing the cause of environmental change from the information, generate explanatory text showing the result of environmental change from the event information of the extraordinary event, and show the causal relationship of environmental change by combining these explanatory text A story sentence is generated.

[Configuration of the embodiment]
Next, the configuration of the story generation system according to the present embodiment will be described with reference to FIG.
The story generation system 10 includes an information processing device such as a server device or a workstation as a whole, detects an event from the sensor network 30, and sequentially stores event information including occurrence status data related to the event. And a story generation processing unit 10B that generates a story sentence indicating a causal relationship of an environmental change corresponding to a newly detected event based on event information of various events accumulated in the event processing unit 10A. Has been.

The event processing unit 10A includes, as main function units, an event detection unit 11, an event database (hereinafter referred to as an event DB) 12, an extraordinary event determination unit 13, and an extraordinary event database (hereinafter referred to as an extraordinary event DB) 14. Is provided.
The story generation processing unit 10B includes, as main function units, a question generation unit 21, a question queue 22, a question processing unit 23, a user interface unit (hereinafter referred to as a user I / F unit) unit 24, an explanatory sentence database ( Hereinafter, a description DB 25) and a story generation unit 26 are provided.

  The event detection unit 11 has a function of acquiring, from the sensor network 30, occurrence status data related to detection events sequentially detected in time series by a sensor of the sensor network 30, and the acquired occurrence status data of the detection event. Along with text information indicating the contents and time information indicating the detection time of the occurrence status data, it has a function of accumulating as event information in the event DB 12.

The time window is a time section having a predetermined time length, and is continuously arranged at the same time position with a period of day, week, month, or the like. For this reason, events detected in the past are stored in association with time windows as event information. In addition, not only a time window having one type of time length but also a plurality of time windows having different time lengths may be set as time windows. Moreover, you may use the time window of a different time length with the event and sensor used as a detection target.
In the time window, a plurality of sensor detection timings are set at equal intervals, and the occurrence status data is sequentially detected in time series for each detection timing. For this reason, each occurrence status data obtained in one time window indicates a time-series change regarding the occurrence status of the event detected by the sensor.

  Occurrence status data includes data indicating the occurrence and number of events related to the movement and movement of objects detected by motion sensors such as acceleration sensors and magnetic orientation sensors attached to objects in the real environment, as well as events. Data indicating physical quantities such as temperature, humidity, brightness, sound, power consumption, etc. may be used. The value of the occurrence status data may be an instantaneous value at the detection timing, or a statistical value such as an average value, a representative value, or a cumulative value regarding data detected in a plurality of detection periods between the immediately preceding detection timings.

  Therefore, the detected events include from advanced events such as “tea brew” and “cleaning” to low-level events such as “cups move”. An advanced event can be detected by a method using supervised learning (see, for example, Non-Patent Document 6). Further, low-level events such as the movement of an object can be detected by a method (for example, see Non-Patent Document 7) that detects a time interval in which the acceleration data of the acceleration sensor changes greatly.

  The explanation information included in the event information is text information explaining the name and content of the event. Since the sensor identification information indicating the sensor that has detected the occurrence status data is notified together with the occurrence status data, the event DB 12 and the storage unit (see FIG. 6) are associated with the sensor identification information and the event description information corresponding to the occurrence status data. (Not shown), it is possible to easily obtain the explanation information corresponding to the detection event.

  The event DB 12 includes a storage device such as a hard disk, and has a function of accumulating event information related to the detection event acquired by the event detection unit 11. The event information includes a set of occurrence status data detected at each detection timing within the time window, time information indicating the detection time of these occurrence status data, and explanation information explaining the name and content of the event.

  The extraordinary event determination unit 13 is characterized by time-series changes regarding the occurrence status data of the detected event newly detected by the event detection unit 11, and the occurrence status data of the past event detected in the past same time window acquired from the event DB 12. A function that calculates an index indicating the consistency with the characteristics of time-series changes related to this and whether or not the detected event is an extraordinary event that deviates from the past event according to the comparison result between the index and the threshold It has the function to judge.

  As a method for determining an extraordinary event from past events, a known method for determining the coincidence of two data may be used. In the present embodiment, for event information of past events stored in the event DB 12 and event information of newly detected detection events, outliers based on an information loss rate by dimension compression and a mixed normal distribution for each principal component The probability density for performing detection is calculated as an index for determining an unusual event.

  As a function for determining an extraordinary event based on the information loss rate, the extraordinary event determination unit 13 uses a detection vector having an occurrence state related to detection events sequentially detected in a time series in an arbitrary time window as an element in the same time window. A function to perform principal component analysis processing on a plurality of past vectors related to past events obtained in the past, a function to perform dimension compression processing of detection vectors based on the obtained main principal components, and information loss about the detection vectors A function for calculating the rate, and a function for determining whether or not the detected vector is an extraordinary event by comparing the information loss rate and the loss rate threshold value.

Further, as a function of determining an extraordinary event based on the probability density, the extraordinary event determination unit 13 projects the detection vector and the past vector on the main principal component axis described above, and applies the corresponding for each main principal component axis. By comparing the probability density related to the projection value of the detection vector in the mixed normal distribution and the probability density threshold for each main principal component axis with the function of generating a mixed normal distribution related to the projection value, And a function for determining whether the event is an extraordinary event.
Specific calculation methods related to the information loss rate and probability density, and the determination order using these indices will be described in detail later.

  The extraordinary event DB 14 includes a storage device such as a hard disk, and has a function of accumulating event information related to extraordinary events determined by the extraordinary event determination unit 13. The event information stored in the extraordinary event DB 14 has the same configuration as the event information stored in the event DB 12.

The question generation unit 21 selects a past event correlated with the detected event determined as the extraordinary event by the extraordinary event determination unit 13 from the event DB 12 and the extraordinary event DB 14, and event information of the selected past event And a function for generating question data for inquiring about the cause of occurrence of the extraordinary event using these past events as answer candidates, and a function for sequentially storing these question data in the question queue 22. .
As a method for selecting a past event correlated with the detected event, a known method such as an a priori algorithm may be used. The past event selection method using the a priori algorithm will be described in detail later.

  The question queue 22 includes a storage device such as a semiconductor memory or a hard disk, sequentially stores the question data generated by the question generation unit 21, and outputs the question data in the order in which the question data is stored in response to reading from the question processing unit 23. It has a function to do.

  The question processing unit 23 reads the question data from the question queue 22 in response to the user detection notification from the user I / F unit 24, and presents the content of the question data to the user via the user I / F unit 24. And a function for identifying a cause event that causes a detection event determined as an extraordinary event based on a response from the user to the presentation received via the user I / F unit 24, and the identified cause An explanatory sentence DB25 that combines the generated explanatory text with a function that generates an explanatory text showing the cause of the environmental change from the event information of the event, and an explanatory text showing the result of the environmental change from the event information of the detected event And has a function of saving data.

  The explanatory note DB 25 includes a storage device such as a semiconductor memory or a hard disk, and has a function of storing a set of an explanatory note relating to the cause event generated by the question processing unit 23 and an explanatory note relating to the result event.

  The story generation unit 26 reads a set of explanatory texts generated by the question processing unit 23 from the explanatory text DB 25 and combines the explanatory texts to generate a story sentence that indicates the causal relationship of environmental changes, The posting instruction including the story sentence is transmitted to the Weblog system 41 via the communication network 40.

[Operation of this embodiment]
Next, the operation of the story generation system according to the present embodiment will be described with reference to FIG. FIG. 3 is a flowchart showing the story generation operation of the story generation system according to the present embodiment.

  First, the event detection unit 11 acquires occurrence status data related to detection events sequentially detected in time series by an sensor of the sensor network 30 from the sensor network 30 (step 100). Along with text information indicating the contents of the detected event and time information indicating the detection time of occurrence status data, the event information is stored in the event DB 12 as event information (step 101).

  When the event detection unit 11 acquires a new detection event and accumulates it in the event DB 12, the extraordinary event determination unit 13 reads out the detection event, characteristics of the time series change regarding the occurrence status data of the detection event, Information loss rate calculation processing (step 102) and probability density calculation in order to calculate an index indicating the coincidence with the characteristics of the time-series change relating to the occurrence status data of the past event detected in the same past time window acquired from the DB 12 The process (step 103) is executed.

Thereafter, the extraordinary event determination unit 13 performs an extraordinary event determination process for determining whether or not the detected event is an extraordinary event deviating from the past event according to the comparison result between the index and the threshold. Execute (Step 104)
Thereby, when it is determined that the detected event is an unusual event, the unusual event determining unit 13 copies the event information of the detected event from the event DB 12 to the unusual event DB 14.

  When the extraordinary event determination unit 13 accumulates a new detected event as an extraordinary event in the extraordinary event DB 14, the question generating unit 21 uses the past event correlated with the detected event as an answer candidate. A question generation process for generating question data for inquiring about the cause of occurrence of the problem is executed (step 105), and the generated question data is stored in the question queue 22.

  The question processing unit 23 reads the question data from the question queue 22 in response to the user detection in the user I / F unit 24 by the detection unit such as a bed pressure sensor or an operation button, and presents it to the user by a screen display or a voice message. Then, based on the user's response to this presentation, the cause event that causes the detection event determined as an extraordinary event is identified, and the question processing that generates the cause event and the description of the detection event is executed. (Step 106), the generated set of explanatory texts is stored in the explanatory text DB 25.

Thereafter, the story generation unit 26 reads out a set of explanatory sentences generated by the question processing unit 23 from the explanatory sentence DB 25, and generates a story sentence indicating the causal relationship of the environmental change by combining these explanatory sentences ( Step 107), a posting instruction including the generated story sentence is transmitted to the Weblog system 41 via the communication network 40, and the series of story generation operations is terminated.
Thereby, in the Weblog system 41, in response to a posting instruction from the story generation system 10, the story sentence included in the posting instruction is posted to the blog and released on the communication network 40.

[Information loss rate calculation processing]
Next, an information loss rate calculation process executed by the extraordinary event determination unit 13 will be described with reference to FIGS. 4 and 5. FIG. 4 is a flowchart showing information loss rate calculation processing. FIG. 5 is an explanatory diagram showing a part of the information loss rate calculation process.
The extraordinary event determination unit 13 executes the information loss rate calculation process of FIG. 4 in step 102 of FIG.

First, the extraordinary event determination unit 13 acquires from the event DB 12 event information relating to a newly detected event detected in the time window W, and generates a detection vector V from the occurrence state data included in the event information (step) 110).
As shown in FIG. 5, the time window W is provided with N detection periods (N is an integer equal to or greater than 2) having a time width c. In this example, an event occurrence state for each detection period. , DN are detected and stored as event information.
The extraordinary event determination unit 13 generates an N-dimensional detection vector V having these N occurrence status data d1, d2,..., DN as elements.

Further, the extraordinary event determination unit 13 acquires event information of past events detected in the same time window W from the event DB 12, and generates past vectors corresponding to these past events in the same manner as described above ( Step 111). In the example of FIG. 5, L pieces of N-dimensional past vectors q1, q2,..., QL (L is an integer of 1 or more) are generated.
Thereafter, the extraordinary event determination unit 13 combines these detection vectors V and past vectors q1, q2,..., QL, and analyzes M × N dimensions (M = L + 1) consisting of vectors x1, x2,. A target matrix X is generated (step 112).

  Subsequently, the extraordinary event determination unit 13 performs a singular value decomposition process (SVD: Singular Value Decomposition) on the analysis target matrix X, and calculates a principal component matrix T related to the analysis target matrix X (step 113). As a result, as shown in FIG. 5, a K × M-dimensional (K is an integer of 2 or more) principal component matrix Z composed of principal component vectors z1, z2,..., ZH,. Note that several algorithms have been proposed for the singular value decomposition processing, and these known algorithms may be used.

  Next, the extraordinary event determination unit 13 selects the principal component vectors z1, z2,..., ZH,..., ZK included in the principal component matrix Z based on the contribution rate and factor loading obtained by the singular value decomposition process. .., ZH are extracted as main principal component matrix U (step 114). At this time, the number of principal component vectors to be extracted may be any number (H is an integer of 2 ≦ H <K), but empirically, by setting H = K / 2, it is possible to determine an unusual event well. did it.

Thereafter, the extraordinary event determination unit 13 projects the detection vector V onto the main principal component axis composed of the obtained main principal component matrix U to generate a projection vector V ′ (step 115). The information loss rate R is calculated from the vector V and the projection vector V ′ using the following equation (1) (step 116). In Expression (1), || V || represents the norm of the vector V, and <V, V ′> represents the inner product of the vector V and the vector V ′.

This information loss rate R is obtained from the detection vector V when the dimensions of the principal component of the analysis matrix X including the detection vector V and the past vectors q1, q2,..., QL are compressed from the original K dimension to the H dimension. It shows the rate of information loss.
Therefore, when the amount of information loss is large, the detection vector V includes many H-dimensional main principal components, that is, principal components not included in the past vector. For this reason, the event corresponding to the detection vector can be estimated to be an extraordinary event including a main component different from the daily event corresponding to the past vector.

[Probability density calculation processing]
Next, the probability density calculation process executed by the extraordinary event determination unit 13 will be described with reference to FIG. FIG. 6 is a flowchart showing the probability density calculation process.
The information loss rate obtained by the dimensional compression described above is an index indicating the characteristics of the detection vector and the past vector on the time axis, so even if the detection vector and the past vector are distributed on the same principal component axis, it is detected. The norm of the vector and the norm of the past vector may be greatly different.

In this embodiment, in order to discriminate a detection vector having a norm different from the past vector as an extraordinary event, a probability density related to the detection vector is calculated for each principal component, and outlier detection is performed using a mixed normal distribution.
The extraordinary event determination unit 13 executes the probability density calculation process of FIG. 6 in step 103 of FIG.

  First, the extraordinary event determination unit 13 projects the past vectors q1, q2,..., QL onto the main principal component axis of the main principal component matrix U obtained by the information loss rate calculation processing of FIG. , P2,..., PL (step 120), the space formed by the main principal components is divided in the manner of a binary tree, and the projection vector V of the detection vector V obtained by the information loss rate calculation processing of FIG. Region A including 'is selected (step 121).

  Next, the extraordinary event determination unit 13 generates a mixed normal distribution related to the main principal component axis components of the projection vectors p1, p2,..., PL in the region A for each main principal component axis (step 122). For each principal component axis, the probability density regarding the principal principal component axis component of the projection vector V ′ in the mixed normal distribution is calculated (step 123).

The probability density r takes a maximum value when the principal principal component axis component of the projection vector V ′ is located near the center of the mixed normal distribution, and the value decreases as the mixture moves toward the base of the mixed normal distribution.
Therefore, when the probability density r of the projection vector V ′ on any of the principal component axes shows a small value, the projection vector V ′ is the majority of past vectors constituting the mixed normal distribution. You will have a different norm. For this reason, it can be estimated that the detected event corresponding to the projection vector V ′ is an unusual event different from the everyday event corresponding to the past vector.

[Unusual event judgment processing]
Next, an extraordinary event determination process executed by the extraordinary event determination unit 13 will be described with reference to FIG. FIG. 7 is a flowchart showing an extraordinary event determination process.
The extraordinary event determination unit 13 executes the extraordinary event determination process of FIG. 7 in step 104 of FIG.

First, the extraordinary event determination unit 13 compares the information loss rate R related to the detected event calculated in the information loss rate calculation process of FIG. 4 with a preset loss rate threshold (step 130). .
Here, when the information loss rate R is larger than the loss rate threshold (step 130: YES), the extraordinary event determination unit 13 determines the detected event as an extraordinary event (step 133), and The event information is copied from the event DB 12 to the extraordinary event DB 14 (step 134), and the series of extraordinary event determination processing is terminated.

On the other hand, when the information loss rate R is equal to or less than the loss rate threshold (step 130: NO), the extraordinary event determination unit 13 calculates each main principal component axis related to the detected event calculated by the probability density calculation process of FIG. Each probability density is compared with a preset probability density threshold (step 131).
Here, when the probability density regarding any of the main principal component axes is smaller than the probability density threshold (step 131: YES), the unusual event determination unit 13 determines the detected event as the unusual event. (Step 133), the event information of the detected event is copied from the event DB 12 to the extraordinary event DB 14 (Step 134), and the series of extraordinary event determination processing is terminated.

  If the probability density for all the main principal component axes is larger than the probability density threshold (step 131: NO), the unusual event determination unit 13 determines the detected event as an everyday event (step 132). The extraordinary event determination process is terminated.

[Question generation process]
Next, the question generation processing in the question generation unit 21 will be described with reference to FIG. FIG. 8 is a flowchart showing the question generation process.
The question generation unit 21 executes the question generation process of FIG. 8 in step 105 of FIG.

First, the question generation unit 21 searches the event DB 12 and the extraordinary event DB 14 and selects a past event correlated with the extraordinary event acquired from the extraordinary event DB 14 as a cause event for a result event including the extraordinary event. (Step 140).
As a method for calculating the correlation between events, a known method such as an apriori algorithm (Apriori Algoritm) may be used.

FIG. 9 is an explanatory diagram illustrating a certainty factor calculation method used in the a priori algorithm. Here, it is assumed that the two events A and B are input X and Y, and that the three elements “thing”, “action”, and “owner” are the same in both events. Is set as a condition Z to apply.
For example, if the event A has the content “Mr. A's slippers are MOVE”, “Mono” indicates “Slippers”, “Action” indicates “MOVE”, and “Owner” indicates “Mr. A”. Show. Therefore, the method of calculating the certainty factor is applied only when the event B is “Mr. A's slipper is MOVE”.

  Also, in FIG. 9, a certainty factor calculation method used in the a priori algorithm is defined. Here, the time interval T is defined as the time interval T from the point in time that is traced back by the X time window width from the X start time. And, it is defined that “the ratio of T including Y that is Z out of T including X that is Z” is defined as the certainty factor. Thereby, in the time interval T, the frequency with which Y which is Z occurs is calculated as the certainty factor. Further, when X and Y are equal, it is defined that the certainty factor is “1”.

In this embodiment, since an event is a combination of an object and an action for the object, there are three types of correlation elements: correlation between things, correlation between actions, and correlation between owners of an object. Thought. At this time, since there is a thing in which a clear owner does not exist, when the correlation between owners is obtained, the correlation between owners is set to 0 when no owner exists on the other hand. The correlation between the result event and the cause event is calculated based on the following equation (2) from these three types of correlations.

  In this manner, the question generation unit 21 selects a predetermined number of past events correlated with an unusual event in order of strong correlation as candidate causal events, and then selects the past event and the unusual event. Question data including information is generated (step 141), and a series of question generation processing ends.

[Question processing]
Next, the question processing in the question processing unit 23 will be described with reference to FIG. FIG. 10 is a flowchart showing the question process.
The question processing unit 23 executes the question processing of FIG. 10 in step 106 of FIG.

  First, the question processing unit 23 obtains question data from the question queue 22 (step 150), and uses the occurrence status data of the unusual event included in the event information of the unusual event as the user I / F unit 24. To the user (step 151), and the user I / F unit 24 presents the occurrence status data of the cause event candidate included in the event information of the cause event candidate to the user (step 151). 152).

FIG. 11 is an example of a question data presentation screen. Here, a question message for inquiring the cause of the extraordinary event is displayed in the column (a). Further, the occurrence status data of the extraordinary event is displayed as a graph in the column (b), and the occurrence status data of the cause event candidate is also displayed as a graph in the column (c). In the column (d), a selection menu for selecting the name of the user answering the question and an answer operation button are displayed. In the column (e), the cause event candidate selected in the column (c) is displayed. An operation button for exchanging with the extraordinary event of (b) is displayed.
If the “OK” button in the column (d) is pressed, the question processing unit 23 is notified of a reply that the cause event candidate displayed in the column (b) has been identified as the cause event.

In this way, the question processing unit 23 receives the user's answer via the user I / F unit 24, and identifies the cause event for the unusual event based on the answer (step 153).
Thereafter, the question processing unit 23 inserts the event name and event occurrence time included in each event information into a template for explanatory text set in advance for the identified cause event and extraordinary event. Thus, an explanatory text explaining the contents of each event is generated (step 154), and the series of question processing is terminated.

As a result, a set of these explanatory texts generated by the question processing unit 23 is read to the story generating unit 26 via the explanatory text DB 25, and these explanatory texts are combined to show a story sentence indicating the causal relationship of the environmental change. Is generated.
FIG. 12 is an example of generating a story sentence. Here, as a result event, it is described whether the event name is “cooking rice” and the event occurrence period is “2009/03/06 11: 00-12: 00”, and the cause name is the event name Is the “Hinamatsuri Party” and the event occurrence period is “2009/03/06 12: 00-12: 00”.

Therefore, the explanation of the event is generated as the explanation of the result event, “Cooking rice” from “11:00” to “12:00” of “2009/03/06”. An explanation of “Hinamatsuri Party” is generated from “12:00” to “13:00” of “2009/03/06”.
Then, these explanations are inserted into a preset template for the story sentence, and ["2009/03/06" from "11:00" to "12:00" There was an event that said. The reason is that we were preparing for the “Hina Festival Party” from “12:00” to “13:00” on “2009/03/06”. ] Has been generated.

[Effects of the present embodiment]
As described above, according to the present embodiment, the unusual event determination unit 13 uses the time-series change feature regarding the occurrence status data of the detected event newly detected by the event detection unit 11 and the same past time acquired from the event DB 12. Based on the coincidence with the characteristics of the time series change related to the past event occurrence status data detected in the window, it is determined whether or not the detected event is an extraordinary event that deviates from the past event. Event information of detected events is accumulated in the extraordinary event DB 14, and the question generating unit 21 selects past events correlated with the extraordinary event determined by the extraordinary event determining unit 13 from the event DB 12 and the extraordinary event DB 14. Based on the event information of these past events, Question data for inquiring about the cause of the occurrence of the extraordinary event as a candidate is generated, and based on the answer from the user to the inquiry to the user based on the question data generated by the question generation unit 21 by the question processing unit 23 The cause event that causes the extraordinary event is identified, the explanation of the cause of the environmental change is generated from the event information of the cause event, and the result of the environmental change is explained from the event information of the extraordinary event. A sentence is generated, and the story sentence indicating the causal relationship of the environmental change is generated by the story generator 26 by combining the explanatory sentences generated by the question processor 23.

As a result, a story showing the causal relationship of environmental changes with a very simple user task of selecting a cause event for a result event consisting of a newly detected detection event from among answer candidates presented as question data. A sentence can be generated. In addition, the user automatically selects a non-daily event that does not occur in daily life from a vast number of events that occur in daily life, and automatically generates a story sentence that uses this non-daily event as a result event. Can be generated.
Therefore, according to the present embodiment, it is possible to generate a story sentence indicating a causal relationship regarding a useful environmental change among environmental changes that have actually occurred with a small user burden.

  In the present embodiment, the extraordinary event selection unit 13 uses a detection vector having an occurrence state related to detection events sequentially detected in a time series in an arbitrary time window, and a past obtained in the past in the same time window. Principal component analysis processing is performed on a plurality of past vectors related to the event, and by performing dimension compression processing of the detection vector based on the obtained main principal component, an information loss rate for the detection vector is calculated, and this information loss By comparing the rate and the loss rate threshold, it is determined whether or not the detection vector is an extraordinary event. Therefore, it does not appear in the occurrence status data value itself, but is inherent in the occurrence status data. Differences in principal components that indicate features can be captured, and unusual events can be determined with higher accuracy.

  In the present embodiment, when the information loss rate is equal to or less than the loss rate threshold value in the extraordinary event selection unit 13, the detection vector and the past vector are projected onto the main principal component axis, By generating a mixed normal distribution for the projection value for each principal component axis and comparing the probability density for the projection value of the detection vector in the mixed normal distribution for each major principal component axis and the probability density threshold, Since it is determined whether or not the detection vector is an extraordinary event, even when the information loss rate described above is small, it is possible to capture the difference in norm between the detection vector on the principal component axis and the past vector, Unusual events can be determined with higher accuracy.

[Extended embodiment]
The present invention has been described above with reference to the embodiments, but the present invention is not limited to the above embodiments. Various changes that can be understood by those skilled in the art can be made to the configuration and details of the present invention within the scope of the present invention.

  In the present embodiment, an example is given in which each functional unit in the story generation system 10 is divided into an event processing unit 10A and a story generation processing unit 10B according to processing information to be passed between each other and the execution order of the processing. However, the present invention is not limited to this. These functional units may be provided in one device or may be divided into three or more devices.

  DESCRIPTION OF SYMBOLS 10 ... Story generation system, 10A ... Event processing part, 10B ... Story processing part, 11 ... Event detection part, 12 ... Event DB, 13 ... Extraordinary event determination part, 14 ... Extraordinary event DB, 21 ... Question generation part, DESCRIPTION OF SYMBOLS 22 ... Question queue, 23 ... Question processing part, 24 ... User I / F part, 25 ... Explanation sentence DB, 26 ... Story generation part, 30 ... Sensor network, 31 ... Communication line, 40 ... Communication network, 41 ... Weblog system .

Claims (7)

A story generation system that generates a story sentence indicating a causal relationship of an actual environmental change based on a detection event detected by a sensor,
An event detection unit that accumulates occurrence status data related to the detection events sequentially detected by the sensor in time series, together with explanatory information indicating the contents of the detection events, as event information for each time window of a predetermined time length,
Characteristics of time-series change regarding the occurrence status data of the detected event newly detected by the event detection unit, and characteristics of time-series change regarding the occurrence status data of the past event detected in the same past time window acquired from the event database Based on the coincidence of the event, it is determined whether the detected event is an extraordinary event deviating from the past event, and the event information of the detected event determined as an extraordinary event is accumulated in the extraordinary event database. An event determination unit;
A past event correlated with the extraordinary event determined by the extraordinary event determination unit is selected from the event database and the extraordinary event database, and the past event is selected as a candidate based on event information of the past event. And a question generator for generating question data for inquiring about the cause of the extraordinary event,
Based on the answer from the user to the inquiry to the user based on the question data generated by the question generation unit, the cause event that causes the extraordinary event is identified, and the environmental change is determined from the event information of the cause event. A question processing unit for generating an explanatory sentence indicating the cause of the environmental change and generating an explanatory sentence indicating the result of the environmental change from the event information of the extraordinary event;
A story generation system comprising: a story generation unit that generates a story sentence that indicates the causal relationship of the environmental change by combining explanatory texts generated by the question processing unit. The story generation system according to claim 1,
The extraordinary event selection unit includes a detection vector having an occurrence state related to the detected event sequentially detected in a time series in an arbitrary time window, and a plurality of past vectors related to past events obtained in the past in the same time window; Is subjected to a principal component analysis process, and a dimension compression process of the detection vector is performed based on the obtained main principal component to calculate an information loss rate for the detection vector, and the information loss rate and the loss rate threshold are calculated. A story generation system characterized by determining whether or not the detected vector is the extraordinary event by comparing with a value. The story generation system according to claim 2,
When the information loss rate is equal to or less than the loss rate threshold, the extraordinary event selection unit projects the detection vector and the past vector onto the main principal component axis, and And generating a mixed normal distribution related to the projection value and comparing the probability density related to the projection value of the detection vector in the mixed normal distribution and the probability density threshold for each of these principal principal component axes, A story generation system that determines whether or not the event is an extraordinary event. A story generation method used in a story generation system that generates a story sentence indicating a causal relationship of an actual environmental change based on a detection event detected by a sensor,
The event detection unit accumulates the occurrence status data related to the detected events sequentially detected in time series by the sensor, together with explanatory information indicating the contents of the detected events, as event information for each time window of a predetermined time length in the event database. An event detection step;
The characteristics of the time series change regarding the occurrence status data of the detected event newly detected in the event detection step and the occurrence status data of the past event detected in the past same time window acquired from the event database. On the basis of the coincidence with the characteristics of the time-series change regarding the event, it is determined whether or not the detected event is an extraordinary event deviating from the past event, and the event information of the detected event determined as the extraordinary event is extraordinary Extraordinary event judgment step to accumulate in the event database;
The question processing unit selects a past event correlated with the extraordinary event determined in the extraordinary event determination step from the event database and the extraordinary event database, and based on event information of these past events, A question generation step for generating question data for inquiring about the cause of the occurrence of the extraordinary event, with past events as candidate answers;
The question processing unit identifies a cause event that causes the extraordinary event based on an answer from the user to an inquiry to the user based on the question data generated in the question generation step, and the event of the cause event A question processing step of generating an explanatory sentence indicating the cause of the environmental change from the information and generating an explanatory sentence indicating the result of the environmental change from the event information of the extraordinary event;
A story generation method comprising: a story generation step, wherein a story generation unit generates a story sentence that indicates the causal relationship of the environmental change by combining the explanatory sentences generated in the question processing step. The method for generating a story according to claim 4,
The extraordinary event selection step includes a detection vector having an occurrence state related to the detected event sequentially detected in a time series in an arbitrary time window, and a plurality of past vectors related to past events obtained in the past in the same time window; Is subjected to a principal component analysis process, and a dimension compression process of the detection vector is performed based on the obtained main principal component to calculate an information loss rate for the detection vector, and the information loss rate and the loss rate threshold are calculated. A method for generating a story, comprising a step of determining whether or not the detected vector is the extraordinary event by comparing with a value. The story generation method according to claim 5,
In the extraordinary event selection step, when the information loss rate is equal to or less than the loss rate threshold, the detection vector and the past vector are projected onto the main principal component axis, And generating a mixed normal distribution related to the projection value and comparing the probability density related to the projection value of the detection vector in the mixed normal distribution and the probability density threshold for each of these principal principal component axes, A method for generating a story, comprising the step of determining whether or not the event is an extraordinary event.   The program for functioning a computer as each part which comprises the story production | generation system as described in any one of Claims 1-3.