JP5557338B2 - Prediction program, prediction device, and method for predicting result event from user's action sentence - Google Patents

Description

  The present invention relates to a technique for predicting various results (for example, trouble) from a sentence (text) described by a user using a natural language processing technique.

  Conventionally, there is a technique for calculating the difficulty level of trouble avoidance based on the number of causes of trouble searched from a trouble information database (see, for example, Patent Document 1). According to this technique, the trouble information database stores trouble information including type, occurrence status, cause of occurrence, countermeasures, and the like. This assumes that the trouble is difficult to avoid if there are various causes of trouble. This demonstrates the difficulty of avoiding trouble.

  In addition, there is a technique for detecting a specific expression pattern from a sentence by natural language processing and predicting a trouble occurrence probability based on the expression pattern (see, for example, Patent Document 2). The text may be, for example, a report indicating the status of a business project. From this report, an expression pattern (for example, a word and a dependency relationship) remarkably representing the status of the project is detected. The trouble occurrence probability is stored in advance for each expression pattern, and the trouble occurrence probability related to the project is calculated from the detected expression pattern.

  Furthermore, there is a technique for predicting a trouble in a system in which a plurality of devices are combined (see, for example, Patent Document 3). According to this technology, for each device or part, trouble information is stored in a database in advance, and it is estimated that a corresponding trouble will occur from the device or part used in the device configuration system. Can do.

JP 2007-199809 A JP 2009-230351 A JP 2010-211162 A

“Twitter API Documentation”, [online], [Search January 16, 2011], Internet <URL: http: //apiwiki.twitter.com/w/page/22554679/Twitter-API-Documentation> “Twitter API Specification”, [online], [Search January 16, 2011], Internet <URL: http: //watcher.moe-nifty.com/memo/docs/twitterAPI50.txt> Kentaro Inui, “Natural Language Processing”, Nara Institute of Science and Technology, [online], [Search January 16, 2011], Internet <URL: http://www.cl.ecei.tohoku.ac .jp / ~ inui / papers / 0711Nihongogaku.pdf> Masayuki Asahara, “PrefixSpan-rel-Sequence Pattern Mining Tool”, Nara Institute of Science and Technology, [online], [Search January 16, 2011], Internet <URL: http: //prefixspan-rel.sourceforge .jp / ＞ Naoki Okazaki, Junichi Sakurai, “Abbreviation Extraction Using Conditional Random Fields”, Graduate School of Information Science and Technology, The University of Tokyo, Japan Science and Technology Agency, [online], [Search January 16, 2011], Internet <URL: http: //yans.anlp.jp/symposium/2007/paper/okazaki.pdf>

  According to the techniques described in Patent Documents 1 and 3, it is necessary to provide in advance a database in which trouble information is registered. To construct a database, it is necessary to analyze an individual's behavior in order to predict a trouble caused by the individual's habits. However, there are also unconscious behaviors in individual habits. For this reason, it is necessary to manually construct a database, which is inevitably expensive. Therefore, it is difficult to detect a habit causing a trouble and thereby predict the trouble.

  Moreover, according to the technique described in Patent Document 2, a report report of a project indicating a situation at the time of occurrence of a trouble is required, and a target sentence is specified. In addition, for each word obtained from the expression pattern, a weight of appearance possibility that will be remarkably described only at the time of trouble is given. However, since the weight is given to each word, the fluctuation (difference) in the expression of the word cannot be absorbed.

  Furthermore, according to the techniques described in Patent Documents 1 to 3, the prediction result is limited to “trouble”, that is, a bad result event, and it is not intended to set a good result event as a prediction result. In addition, in any of the technologies, the project and the system configuration are targeted for trouble prediction.

  Accordingly, the present invention provides a prediction program, a prediction device, and a method for predicting a result event that may occur in the future due to a user's habits due to a user's daily behavior event described in a sentence. The purpose is to provide.

According to the present invention, there is provided a prediction program for causing a computer mounted on an apparatus to function so as to predict a result event that will occur in the future for a user due to a user's action event described in a sentence. There,
An action sentence set accumulating means for accumulating the action sentence described by the user and the description time;
Event set extraction means for extracting an event expression from the action sentences stored in the action sentence set storage means, and extracting a set of “events” consisting of a set of the event expression and its description time;
When a plurality of identical event expressions appear within a predetermined period, habit event extracting means for extracting the event as a habit event,
“Feature vector elements” having a large number of event expressions as elements are preset, and feature vector generation means for generating “feature vectors” specifying the presence or absence of corresponding elements for each custom event,
Prediction model storage means for storing in advance a feature vector group learned for each result event from feature vectors based on a large number of behavioral sentences for learning;
A prediction determination unit that determines whether the feature vector output from the feature vector generation unit belongs to a feature vector group for each result event stored in the prediction model storage unit; And causing the computer to function to predict that the resulting event will occur in the future for the user.

According to another embodiment of the prediction program of the present invention,
Learning behavior sentence set storage means for accumulating learning behavior sentences and description times described by a large number of users;
A learning event set extraction means for extracting an event expression from the action sentences stored in the learning action sentence set storage means, and extracting a set of “events” consisting of a combination of the event expression and its description time;
A result event expression storage means for storing result event expressions in advance;
A result event detecting means for detecting a result event expression stored in the result event storing means in the learning action sentence;
A learning habit event extracting means for extracting the event as a habit event when a plurality of the same event expressions appear within a predetermined period in the past from the description time of the detected result event expression,
Learning feature vector generation means for generating a learning feature vector specifying the presence or absence of the corresponding element for each habit event,
It is also preferable that the prediction model storage means causes the computer to function so as to store the learning feature vector for each result event.

  According to another embodiment of the prediction program of the present invention, the prediction determination means is a binary classification method of a support vector machine (Support Vector Machine) or CRF (Conditional), which is integrated with the prediction model storage means. It is also preferable to have the computer function as an engine based on a sequence labeling scheme of Random Fields (conditional random fields).

  According to another embodiment of the prediction program of the present invention, for the habit event extracting means, the expression described in the action sentence is abstracted using an ontology including upper and lower relations, implication relations, etc. Regarding the concept expression, when a plurality of identical event expressions appear within a predetermined period, it is also preferable to cause the computer to function so as to extract these events as a habit event group.

  According to another embodiment of the prediction program of the present invention, it is also preferable that the computer is made to function so as to further include a comment sentence collecting means for acquiring an action sentence from a blog (Weblog) server or a mini blog server.

According to the present invention, a prediction device that predicts a result event that may occur in the future for a user due to a user's action event described in a sentence,
An action sentence set accumulating means for accumulating the action sentence described by the user and the description time;
Event set extraction means for extracting an event expression from the action sentences stored in the action sentence set storage means, and extracting a set of “events” consisting of a set of the event expression and its description time;
When a plurality of identical event expressions appear within a predetermined period, habit event extracting means for extracting the event as a habit event,
“Feature vector elements” having a large number of event expressions as elements are preset, and feature vector generation means for generating “feature vectors” specifying the presence or absence of corresponding elements for each custom event,
Prediction model storage means for storing in advance a feature vector group learned for each result event from feature vectors based on a large number of behavioral sentences for learning;
A prediction determination unit that determines whether the feature vector output from the feature vector generation unit belongs to a feature vector group for each result event stored in the prediction model storage unit; It is characterized by predicting that the result event that will be generated will occur in the future for the user.

According to another embodiment of the prediction device of the present invention,
Learning behavior sentence set storage means for accumulating learning behavior sentences and description times described by a large number of users;
A learning event set extraction means for extracting an event expression from the action sentences stored in the learning action sentence set storage means, and extracting a set of “events” consisting of a combination of the event expression and its description time;
A result event expression storage means for storing result event expressions in advance;
A result event detecting means for detecting a result event expression stored in the result event storing means in the action sentence;
A learning habit event extracting means for extracting the event as a habit event when a plurality of the same event expressions appear within a predetermined period in the past from the description time of the detected result event expression,
Learning feature vector generation means for generating a learning feature vector specifying the presence or absence of the corresponding element for each habit event,
It is also preferable that the prediction model accumulation unit accumulates learning feature vectors for each result event.

  According to another embodiment of the prediction apparatus of the present invention, the prediction determination means is an engine based on a binary classification method of a support vector machine or a CRF sequence labeling method, which is integrated with a prediction model storage means. It is also preferable to make the computer function so that

According to the present invention, there is provided a prediction method using an apparatus so as to predict a result event that will occur in the future due to a user's action event described in a sentence,
An action sentence set accumulating unit for accumulating the action sentence described by the user and the description time;
“Feature vector elements” with a large number of event expressions as elements, and a prediction model in which feature vector groups learned for each result event are pre-stored from feature vectors based on a large number of learning behavior sentences A storage unit,
A first step of extracting an event expression from the action sentences accumulated in the action sentence set accumulating unit, and extracting a set of “events” composed of a combination of the event expression and the description time;
A second step of extracting the event as a habit event when a plurality of identical event expressions appear within a predetermined period;
A third step of generating a “feature vector” designating the presence or absence of the corresponding element for each habit event;
A fourth step of determining whether the feature vector belongs to any of the feature vector groups for each result event stored in the prediction model storage unit, and the result event determined by the prediction determination unit is It is characterized by predicting that it will occur to the user in the future.

  According to the prediction program, the prediction apparatus, and the method of the present invention, a result event that will occur in the future due to the user's habits due to the user's daily behavior event described in the sentence is predicted. be able to.

It is explanatory drawing showing the example of the result event estimated with respect to the action event. It is a functional block diagram of the prediction program in this invention. It is explanatory drawing showing the action sentence accumulate | stored in the action sentence set storage part. It is explanatory drawing showing the event expression extracted from the action sentence of FIG. It is the 1st explanatory view showing extraction of a habit event in the present invention. It is the 2nd explanatory view showing extraction of a habit event in the present invention. It is explanatory drawing showing the production | generation of the feature vector in this invention. It is explanatory drawing showing detection of the result event in this invention. It is explanatory drawing showing the production | generation of the learning feature vector in this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is an explanatory diagram illustrating an example of a result event predicted for a behavior event.

  According to the prediction program of the present invention, the “result event” resulting from the habit of the individual user is predicted from the “behavior event” of the individual user. The action sentence describing the action event can be collected from a site such as a blog or a mini blog (for example, twitter (registered trademark)) connected to the Internet. Such a sentence describes a daily action in an individual user like a diary.

For example, a result event can be predicted from the behavior of an individual user's behavior event as follows.
Action "Eat curry and beef bowl every day"-> Result "High blood pressure"
(Food habits rich in salt and fat)
Action “Smoking cigarettes every day”-> Result “High blood pressure”
Action "Night Stay"-> Result "Late Late"
Action "Insufficient study"-> Result "Follow-up test / Failure"

  Here, the result event predicted by the present invention is limited to a result caused by an individual's behavior habit, such as “high blood pressure”. On the other hand, accidental events such as “earthquake” and “flood” and social events such as “war” and “refugee” are not covered by the present invention.

  Moreover, about this invention, a habitual (periodic) action event is detected as a habit event. Similar habit events are detected as the same habit events conceptually. For example, “eating ramen in front of the station at noon” and “eating beef bowl at noon” are preferably viewed as the same action event as “a diet rich in salt and fat”.

  FIG. 2 is a functional configuration diagram of the prediction program in the present invention.

  According to the prediction program in FIG. 2, the comment sentence collection unit 100, the action sentence set accumulation unit 101, the event set extraction unit 102, the habit event extraction unit 103, the feature vector generation unit 104, and the prediction model accumulation unit 105. And a prediction determination unit 106. These functional units can be realized by executing a program that causes a computer installed in the apparatus to function. The processing flow of these functional components is also understood as a prediction method in the apparatus.

  The comment sentence collection unit 100 acquires a comment sentence of daily activities described by the user from the blog server or the mini blog server. The acquired comment sentence is output to the action sentence information storage unit 101. For example, according to the twitter site server, a comment sentence (tweet) based on a specified user name can be searched. You can also search for tweets for up to one week by specifying the date. These API specifications are described in Non-Patent Documents 1 and 2, for example.

  The action sentence set accumulation unit 101 accumulates the action sentence (comment sentence) and description time of the user collected by the comment sentence collection unit 100.

  FIG. 3 is an explanatory diagram illustrating the action sentences accumulated in the action sentence set accumulation unit.

  According to FIG. 3, the action sentence set accumulation unit 101 accumulates action sentences and description times described by a specific individual user A.

  The event set extraction unit 102 extracts “event expression” from the action sentences accumulated in the action sentence set accumulation unit 101.

  FIG. 4 is an explanatory diagram showing event expressions extracted from the action sentence of FIG.

  According to FIG. 4, the event expression extracted by the event set extraction unit 102 is associated with the description time. The extracted event expressions are sorted in chronological order of description times.

  A sentence consists of a plurality of “words”. A “word” is a structural unit of a language and consists of one or more morphemes. A morpheme is the smallest unit that has meaning for a language, and refers to each of a set of phonemes that have been decomposed into units that no longer make sense after further decomposition. The word includes a simple word (for example, “mountain”) composed of one morpheme and a compound word (for example, “mountain climb”) composed of a plurality of morphemes. Here, when referring to “words”, both simple words and compound words are targeted.

  Phrases, clauses, sentences and sentences are composed by gathering words. This uses a syntax analyzer to extract a pair of a noun and a verb or adjective in a dependency relationship as an “event expression”. For example, a sentence “order a lunch” is composed of three words “bento”, “to” and “order”. The words can be broadly classified into function words having a grammatical role (“O”) and content words having other general meanings (“Bento” and “order”). Here, “event expression” means a sentence including a content word composed of at least a noun and a verb. When there is no verb or adjective, an expression in which nouns are continuous is extracted as an event expression. The event expression is an expression representing a specific event, and its description is not particularly limited.

  Here, according to FIG. 4, correction of the description time is represented. Using the event expression (for example, “10 years ago”, “this year”) or the function word (for example, “was”), the description time is corrected to “event occurrence time (t ′)”.

  For example, the description time of “complete three days ago” is “October 5, 2010 1:00”. Here, based on “3 days ago”, the description time is returned by 3 days and is corrected to the event occurrence time “October 6, 2010, 9:00”.

  Then, the event set extraction unit 102 outputs a set of “events” composed of sets of event expressions and description times to the habit event extraction unit 103.

  The habit event extraction unit 103 inputs a set of event expressions for the user, and when a plurality of identical event expressions appear within a predetermined period, the event is extracted as a “habit event”. The extracted habit event is output to the feature vector generation unit 104.

  FIG. 5 is a first explanatory diagram showing extraction of habit events in the present invention.

  According to FIG. 5, if the same event expression appears for a predetermined period in the past from the description time t for each event expression s, the habit event extracting unit 103 determines that the event is a learning habit event E. Extracted as'. Here, the “predetermined period” is a period having a habit such as 24 hours or 3 days. If there is no coincidence with other event expressions in the past, it is regarded as one event E '

  For example, as seen from “Gyudon @ Ekimae” “October 2, 2010 12:00”, there is an event “Ramen @ Ekimae” “October 1, 2010 19:00” within the past 24 hours. These are extracted as habit events E '.

  Also, for example, “Three days ago,” “October 5, 2010 1:00:00”, “Let ’s stay up” and “October 2, 2010 0:30” within the past 3 days. There is an event. These are extracted as habit events E '.

  FIG. 6 is a second explanatory diagram showing extraction of habit events in the present invention.

  According to FIG. 6, the habit event extracting unit 103 abstracts the expression described in the action sentence using “Ontology” including upper and lower relations and implication relations. If a plurality of identical event expressions appear within a predetermined period of the abstracted high-level concept expression, these events are extracted as habit event groups. A learning habit event E ′ is extracted by comparing the abstracted event expressions. As a result, the shaking of the expression of the action event can be eliminated.

  “Ontology” is defined as “explicit specification of conceptualization” in the field of artificial intelligence technology. Conventionally, according to the text search technique, it is possible to search only by word unit match or similarity. On the other hand, ontology refers to a structure that defines terms describing semantic information (metadata) that describes the content of the sentence. This semantic information is added to the sentence. A sentence can be searched using this semantic information as a search target. By introducing ontology, the sentence to be searched is not treated as a mere collection of words but as data having a large meaning in the whole sentence. According to the present invention, habituality can be detected as a high-level concept expression by abstracting (eventing information) an event expression of an action sentence using an ontology. Further, a paraphrasing technique for natural language processing can be used instead of ontology (see, for example, Non-Patent Document 3).

  In addition, using the pattern detection technology for the abstracted event expression, the event sequence that appears characteristically in the event set is automatically identified, and the resulting event sequence is extracted as the learning habit event E ′. May be. As a pattern detection technique, for example, there is PrefixSpan which is a sequence pattern mining algorithm (see, for example, Non-Patent Document 4).

  The feature vector generation unit 104 inputs a habit event, and generates a “feature vector” that specifies the presence or absence of the corresponding element for each habit event. Here, “feature vector elements” having a large number of event expressions as elements are preset. Then, the feature vector generation unit 104 outputs the generated feature vector to the prediction determination unit 106.

  FIG. 7 is an explanatory diagram showing generation of feature vectors in the present invention.

  According to FIG. 7, a feature vector is generated for each set of habit events E ′. The “feature vector” is an element in which a predetermined event expression is used as an element and the presence or absence is assigned to each element (dimension). The initial value of each element of the feature vector is 0. For each habit event E ', each element corresponding to the event expression is set to 1. Each element of the feature vector is defined in advance as a system.

  According to FIG. 7, for example, in the custom event E ′ of “Ramen @ Ekimae” and “Gyudon @ Ekimae”, the elements of the event expression “4. Ramen Ekimae” and the event expression “9. "

  The prediction model accumulation unit 105 accumulates in advance a feature vector group learned for each result event from feature vectors based on a large number of learning behavior sentences. Further, the prediction model accumulation unit 105 may accumulate a learning feature vector for each result event. The construction of learning feature vectors stored in the prediction model storage unit 105 will be described later.

  The prediction determination unit 106 determines whether the feature vector input from the feature vector generation unit 104 belongs to a feature vector group for each result event stored in the prediction model storage unit 105. The prediction determination unit 106 outputs the result event as an event that is an event that will occur in the future for the user.

  The prediction determination unit 106 is configured integrally with the prediction model storage unit 105, a support vector machine binary classification method, or a CRF (Conditional Random Fields) conditional labeling method. It may be an engine based on

  The “support vector machine” is a linear binary classifier and is used when there are two classes. The two classes are called a positive class and a negative class, respectively. Usually, a class to be analyzed is a positive class, a case belonging to the positive class is called a positive example, and a case belonging to the negative class is called a negative example. Here, the feature vector is a positive example, and the learning feature vector is a negative example. “CRF” is a method in which a logarithmic linear model, which is a probabilistic classifier, is applied to sequence labeling (see, for example, Non-Patent Document 5).

  Next, generation of learning feature vectors accumulated in the prediction model accumulation unit 105 will be described. In the learning stage, in order to predict a result event, a feature vector is extracted from a learning behavior event that appears before the occurrence of the result event. Then, a learning habit event that tends to cause a result event is statistically calculated.

  The prediction program includes a learning behavior sentence set accumulation unit 111, a learning event set extraction unit 112, a learning habit event extraction unit 113, a learning feature vector generation unit 114, a result event expression accumulation unit 121, and a result event detection unit 122. It has further. These functional components can also be realized by executing a program that causes a computer installed in the apparatus to function.

The learning action sentence set accumulation unit 111 accumulates learning action sentences and description times described by a large number of users. The behavioral sentence set accumulation unit 101 accumulates a behavioral sentence set of users to be predicted, while the learning behavioral sentence set accumulation unit 111 accumulates a large number of user behavioral sentence sets. Other than that, the behavior sentence set accumulation unit 101 is exactly the same.
Learning behavior sentence set: D _T
D _T = {d _u1 , d _u2 ,..., D _un }
(D _ui : action sentence of user i, n: number of users)

  The learning event set extraction unit 112 extracts an event expression from the learning action sentences stored in the learning action sentence set storage unit 111, and an “event” composed of a combination of the extracted event expression and its description time. Extract a set. The event set extraction unit 102 extracts a user's event sentence set to be predicted, whereas the learning event set extraction unit 112 extracts an event sentence set for each user for a large number of users. The rest is exactly the same as the event set extraction unit 102.

  The result event expression storage unit 121 stores result event expressions in advance. The result event expression may be a negative trouble term such as “high blood pressure”, “shortness of breath”, “retest”, or a positive term such as “blood pressure drop”, “physical condition improvement”, “promotion”. It may be.

  The result event detection unit 122 detects the result event expression accumulated in the result event accumulation unit 121 among the learning action sentences accumulated in the learning action sentence set accumulation unit 111. The detected result event expression is output to the learning habit event extracting unit 113.

  FIG. 8 is an explanatory diagram showing detection of a result event in the present invention.

  According to FIG. 8, the learning event set E is classified into “occurrence” or “non-occurrence” of the result event. It should be noted that a result event may be detected using an engine based on the binary classification method of the support vector machine or the CRF sequence labeling method based on the surrounding word of the result event expression and its part of speech. .

  The learning habit event extracting unit 113 inputs a set of events and a result event expression. When a plurality of identical event expressions appear within a predetermined period in the past from the description time of the detected event expression, the learning habit event extracting unit 113 extracts the event as a habit event. Note that the learning habit event extracting unit 113 may abstract the learning habit event extracting unit 113 using an ontology including upper and lower relations and implications as in the habit event extracting unit 103 (see FIG. 6). The learning habit event extracting unit 113 is the same as the processing content of the habit event extracting unit 103. The extracted habit event is output to the learning feature vector generation unit 114.

  The learning feature vector generation unit 114 generates a learning feature vector that specifies the presence or absence of a corresponding element for each habit event. The learning feature vector generation unit 114 is the same as the processing content of the feature vector generation unit 104 (see FIG. 7). The learning feature vector generation unit 114 outputs the learning feature vector to the prediction model accumulation unit 105. The prediction model accumulation unit 105 accumulates learning feature vectors for each result event.

  FIG. 9 is an explanatory diagram showing generation of learning feature vectors in the present invention.

  The learning habit event E ′ that occurs habitually in the past from the result event expression is considered to be possibly caused by the result event. In the learning habit event, the value of the element (dimension) vector is set to 1 for the learning habit event before the threshold time before the occurrence time of the result event, and the others are set to 0.

  The generated learning feature vector may be weighted for each vector. An engine based on the binary classification method of the support vector machine or the sequence labeling method of the CRF can also be used for the weighting. For example, according to the importance of the result event, a high weight can be given to the learning habit event before the threshold time when the result event occurs.

  Instead of weighting each word, weighting each habit event can absorb fluctuations in the expression of the word.

  When a support vector machine is used, feature vectors for learning habit events that occurred before a threshold time before a result event may be combined and used as an input. In addition, when the CRF sequence labeling method is used, a feature vector and a set of label pairs are input, and the result of calculating the weight for each feature vector based on the feature vectors related to the habit event before and after is output. May be.

  As described above in detail, according to the prediction program, the prediction apparatus, and the method of the present invention, due to the daily action event of the user described in the sentence, in the future due to the user's habits. The resulting events that will occur can be predicted.

  The present invention is applicable to a health management system and an examination support system, and can improve a user's lifestyle by predicting a result (failure or success) caused by an individual's habit.

  According to the present invention, a result event (for example, a trouble) can be predicted based on a comment sentence such as a diary or a blog that an individual describes on a daily basis without constructing a dedicated database. That is, a sentence that clearly describes the situation in which the result event has occurred is not required.

  Various changes, modifications, and omissions of the above-described various embodiments of the present invention can be easily made by those skilled in the art. The above description is merely an example, and is not intended to be restrictive. The invention is limited only as defined in the following claims and the equivalents thereto.

DESCRIPTION OF SYMBOLS 100 Comment sentence collection part 101 Action sentence set accumulation part 102 Event set extraction part 103 Habit event extraction part 104 Feature vector generation part 105 Prediction model accumulation part 106 Prediction judgment part 111 Learning action sentence set accumulation part 112 Learning event set extraction part 113 Learning Habit event extraction unit 114 Learning feature vector generation unit 121 Result event expression storage unit 122 Result event detection unit 2 Miniblog server

Claims (9)

A prediction program that causes a computer mounted on a device to function so as to predict a result event that will occur in the future to the user due to a user behavior event described in a sentence,
An action sentence set accumulating means for accumulating the action sentence described by the user and the description time;
Event set extraction means for extracting an event expression from the action sentences stored in the action sentence set storage means, and extracting a set of “events” consisting of a set of the event expression and its description time;
When a plurality of identical event expressions appear within a predetermined period, habit event extracting means for extracting the event as a habit event,
A “feature vector element” having a large number of event expressions as elements is preset, and for each habit event, a feature vector generating means for generating a “feature vector” designating the presence or absence of a corresponding element;
Prediction model storage means for storing in advance a feature vector group learned for each result event from feature vectors based on a large number of behavioral sentences for learning;
Prediction prediction means for determining whether the feature vector output from the feature vector generation means belongs to any of the feature vector groups for each result event stored in the prediction model storage means; A prediction program for causing a computer to function so as to predict that a result event determined by the determination means will occur in the future for the user. Learning behavior sentence set storage means for accumulating learning behavior sentences and description times described by a large number of users;
Learning event set extraction means for extracting an event expression from the action sentences stored in the learning action sentence set storage means, and extracting a set of “events” consisting of a set of the event expression and its description time;
A result event expression storage means for storing result event expressions in advance;
A result event detecting means for detecting a result event expression stored in the result event storing means in the learning action sentence;
A learning habit event extracting means for extracting the event as a habit event when a plurality of the same event expressions appear within a predetermined period in the past from the description time of the detected result event expression,
Learning feature vector generation means for generating a learning feature vector designating the presence or absence of a corresponding element for each habit event,
The prediction program according to claim 1, wherein the computer is caused to function so that the prediction model storage unit stores the learning feature vector for each result event.   The prediction judging means is a binary classification method of Support Vector Machine or a CRF (Conditional Random Fields) series labeling method, which is integrated with the prediction model storage means. The prediction program according to claim 1 or 2, wherein the computer functions so as to be an engine based on the above.   About the habit event extracting means, the expression described in the action sentence is abstracted using an ontology including upper and lower relations, implication relations, etc., and a plurality of the same upper concept expressions are abstracted within a predetermined period. 4. The prediction program according to claim 1, wherein when an event expression appears, the computer is caused to function so as to extract the events as a habit event group. 5.   5. The prediction program according to claim 1, further comprising a comment sentence collecting unit configured to acquire the action sentence from a blog (Weblog) server or a mini blog server. . A prediction device for predicting a result event that may occur in the future for a user due to a user action event described in a sentence,
An action sentence set accumulating means for accumulating the action sentence described by the user and the description time;
Event set extraction means for extracting an event expression from the action sentences stored in the action sentence set storage means, and extracting a set of “events” consisting of a set of the event expression and its description time;
When a plurality of identical event expressions appear within a predetermined period, habit event extracting means for extracting the event as a habit event,
A “feature vector element” having a large number of event expressions as elements is preset, and for each habit event, a feature vector generating means for generating a “feature vector” designating the presence or absence of a corresponding element;
Prediction model storage means for storing in advance a feature vector group learned for each result event from feature vectors based on a large number of behavioral sentences for learning;
Prediction prediction means for determining whether the feature vector output from the feature vector generation means belongs to any of the feature vector groups for each result event stored in the prediction model storage means; A prediction apparatus for predicting that a result event determined by the determination means will occur in the future for the user. Learning behavior sentence set storage means for accumulating learning behavior sentences and description times described by a large number of users;
Learning event set extraction means for extracting an event expression from the action sentences stored in the learning action sentence set storage means, and extracting a set of “events” consisting of a set of the event expression and its description time;
A result event expression storage means for storing result event expressions in advance;
A result event detecting means for detecting a result event expression stored in the result event storing means in the action sentence;
A learning habit event extracting means for extracting the event as a habit event when a plurality of the same event expressions appear within a predetermined period in the past from the description time of the detected result event expression,
Learning feature vector generation means for generating a learning feature vector designating the presence or absence of a corresponding element for each habit event,
The prediction apparatus according to claim 6, wherein the prediction feature storage unit stores the learning feature vector for each result event. A prediction method using a device to predict a result event that will occur in the future for the user due to a user behavior event described in a sentence,
An action sentence set accumulating unit for accumulating the action sentence described by the user and the description time;
“Feature vector elements” with a large number of event expressions as elements, and a prediction model in which feature vector groups learned for each result event are pre-stored from feature vectors based on a large number of learning behavior sentences A storage unit,
A first step of extracting an event expression from the action sentences stored in the action sentence set storage unit, and extracting a set of “events” composed of a set of the event expression and its description time;
A second step of extracting the event as a habit event when a plurality of identical event expressions appear within a predetermined period;
A third step of generating a “feature vector” designating the presence or absence of a corresponding element for each habit event;
And a fourth step of determining whether the feature vector belongs to any of the feature vector groups for each result event stored in the prediction model storage unit, and the result event determined by the prediction determination unit Predicting that the user will occur in the future. A learning behavior sentence set accumulation unit that accumulates learning behavior sentences and description times described by a large number of users;
A result event expression storage unit that stores result event expressions in advance;
Extracting an event expression from the action sentences stored in the learning action sentence set accumulating means, and extracting a set of “events” consisting of a set of the event expression and its description time;
Detecting a result event expression stored in the result event storage means in the action sentence;
If a plurality of identical event expressions appear within a predetermined period in the past from the description time of the detected result event expression, extracting the event as a habit event; and
Generating a learning feature vector that specifies the presence or absence of a corresponding element for each habit event,
The prediction method according to claim 8, wherein the learning feature vector is stored in the prediction model storage unit for each of the result events.

Images