Probabilistic retrospective event detection

• US 20070038653A1
• Filed: 10/21/2005
• Published: 02/15/2007
• Est. Priority Date: 08/10/2005
• Status: Active Grant

• Current Assignee: Microsoft Technology Licensing LLC (Microsoft Corporation)
• Original Assignee: Microsoft Corporation
• Inventors: Ma, Wei-Ying, Li, Mingjing, Li, Zhiwei
• Granted Patent: US 7,788,263 B2
• Time in Patent Office: Days
• US Class Current: 1/1
• CPC Class Codes: G06F 16/3344 using natural language analysis
• CPC Class Codes: G06F 16/3346 using probabilistic model

Abstract

Probabilistic retrospective event detection is described. In one aspect, event parameters are initialized to identify a number of events from a corpus of documents. Using a generative model, documents are determined to be associated with an event to detect representative events from the identified number of events.

Claims

1. A computer-implemented method comprising:

initializing event parameters to identify a number of events from a corpus of documents; and

probabilistically determining, using a generative model, whether documents are associated with an event to detect representative events of the number of events.

2. The method of claim 1, wherein the generative model determines probabilistic distributions of persons, locations, keywords, and time conditioned on events to cluster document(s) to events.

3. The method of claim 1, wherein the generative model models document content and timestamps associated with events with different Bayesian mixture models.

4. The method of claim 1, wherein initializing the event parameters further comprises identifying salient events.

5. The method of claim 1, wherein initializing the event parameters further comprises:

selecting an event reported by one or more of the documents; and

for each entity associated with the event;

generating a respective news article for the event; and

determining a time for the respective news article.

6. The method of claim 5, wherein operations associated with the selecting, generating, and determining use one or more of mixing proportions of events, priors of events, conditional multinomial distributions given the event, and parameters of a conditional Gaussian distribution given the event.

7. The method of claim 1, wherein document content associated with an event of the respective events comprises one or more of information associated with persons, locations, and keywords and wherein probabilistically determining further comprises:

modeling the document content with mixtures of different respective unigram models; and

modeling the timestamps with a mixture model that is not one of the different respective unigram models.

8. The method of claim 1, wherein probabilistically determining further comprises:

estimating parameters for the generative model using the event parameters;

generating event clusters to cluster events reported by the documents using estimated generative model parameters;

for each event cluster;

increasing or decreasing a number of events to represent a corresponding event;

if the number of events is not a minimum or a maximum number of events;

(a) again performing operations associated with initializing the event parameters to generate re-initialized event parameters; and

(b) using the generative model to probabilistic detect events from salient ones of the documents using the re-initialized event parameters; and

if the number of events is a minimum or a maximum number of events, summarizing event(s) associated with the event cluster to assign content of one or more documents to respective events.

9. The method of claim 8, wherein estimating the parameters further comprises estimating model parameters using expectation maximization to iteratively maximize log-likelihood.

10. The method of claim 8, wherein generating the event clusters further comprises:

representing probability associated with time as a function of whether a date of the document is associated with a beginning stage of the event or associated with an end stage of the event; and

wherein the beginning stage and the end stage represent a continuous duration of the event.

11. The method of claim 8, wherein generating the event clusters further comprises:

independently estimating, for respective person, location, and keyword models, model parameters by iteratively determining expectation and maximization probabilities, the model parameters being a number of documents and a number of events; and

calculating the probability using the models in view of temporal continuity of the event and any overlap of two or more of the events.

12. A computer-implemented method comprising:

initializing event parameters to identify a salient number of events from a corpus of documents;

estimating parameters for a generative model for probabilistic retrospective detection of events from the salient number of events, the generative model comprising respective models for person(s), time(s), location(s), and keyword(s);

clustering events represented by documents using the parameters for the generative model;

increasing or decreasing a number of events associated with respective ones of clustered events to re-initialize events;

for respective event clusters, if a minimum or maximum number of events has not been reached, again performing operations of the estimating, clustering, and increasing or decreasing; and

for respective event clusters, if a minimum or maximum number of events has been reached, summarizing events in resulting event clusters.

13. The method of claim 12, wherein initializing the event parameters further comprises:

selecting an event reported by one or more documents; and

for each entity associated with the event;

generating a respective news article for the event; and

determining a time for the respective news article.

14. The method of claim 13, wherein operations associated with the selecting, generating, and determining use one or more of mixing proportions of events, priors of events, conditional multinomial distributions given the event, and parameters of a conditional Gaussian distribution given the event.

15. The method of claim 12, wherein the generative model models document content and timestamps associated with events with different Bayesian models.

16. The method of claim 12, wherein the generative model models document content and timestamps associated with events with different Bayesian models, and wherein the document content comprises one or more of information associated with persons, locations, and keywords and wherein estimating the parameters further comprises:

modeling the persons, locations, and keywords with mixtures of different respective unigram models.

17. The method of claim 12, wherein estimating the parameters further comprises modeling time with a particular mixture model that is different than respective mixture models used to model document content associated with an event.

18. The method of claim 12, wherein estimating the parameters further comprises estimating model parameters using expectation maximization to iteratively maximize log-likelihood.

19. The method of claim 12, wherein generating the event clusters further comprises:

independently estimating, for respective person, location, and keyword models, model parameters by iteratively determining expectation and maximization probabilities, the model parameters being a number of documents and a number of events; and

calculating the probability using the models in view of temporal continuity of the event and any overlap of two or more of the events.

20. A computing device comprising processing means for:

setting event parameters to identify documents comprising respective events;

probabilistically detecting events from documents using a multi-modal generative model, the generative model comprising independent mixture models to model document content associated with an event and time associated with an event, the document content comprising information corresponding to one or more of persons, locations, and keywords; and

wherein the processing means iteratively implement operations for setting the event parameters and probabilistically detecting the events until a configurable minimum or maximum number of events associated with respective ones of one or more salient events has been detected.