Method for summarizing event-related texts to answer search queries
A method and apparatus for receiving training data that comprise a plurality of event-and-time-specific texts that are contextually related to a plurality of events; iteratively processing the training data to generate a modified network model that defines a plurality of states; receiving additional data that comprise a plurality of additional event-and-time-specific texts that are contextually related to a particular event; processing the additional data by applying the modified network model to the additional data to identify, within the plurality of additional event-and-time specific texts, a particular set of texts that belong to a particular state of the plurality of states; identifying, within the particular set of texts, one or more texts that are most representative of all texts in the particular set of texts that belong to the particular state; wherein the method is performed by one or more special-purpose computing devices.
1. A method comprising:
receiving training data that comprise a plurality of event-and-time-specific texts that are contextually related to a plurality of events;
identifying, by iteratively processing the training data, a plurality of different states, each state corresponding to a different event from the plurality of events;
generating a modified network model that comprises the plurality of states;
receiving additional data that comprise a plurality of additional event-and-time-specific texts that are contextually related to a particular event;
processing the additional data by applying the modified network model to the additional data to identify, within the plurality of additional event-and-time specific texts, a particular set of texts that belong to a particular state of the plurality of states; and
identifying, within the particular set of texts, one or more texts that are most representative of all texts in the particular set of texts that belong to the particular state; and
wherein the method is performed by one or more computing devices.
2. The method of claim 1 , further comprising:
generating a summary of an event that corresponds to the particular state, wherein the summary is based on the one or more texts that are most representative of all texts in the particular set of texts that belong to the particular state.
3. The method of claim 1 , wherein the modified network model is any one of: a modified Hidden Markov Model (HMM) or a modified Dynamic Bayesian Networks Model (DBNM).
4. The method of claim 1 , further comprising:
initializing one or more elements of a network model;
wherein iteratively processing the training data to generate the modified network model comprises determining one or more parameters of the network model;
identifying a plurality of state-and-event-specific words for the event in a content of the plurality of event-and-time-specific texts; and
including the plurality of state-and-event-specific words in determining the plurality of states within the event and iterating executing of the modified network model using the content of the plurality of event-and-time specific texts.
5. The method of claim 1 , further comprising:
determining unknown parameters of the modified network model using a Baum-Welch algorithm; and
wherein the Baum-Welch algorithm computes maximum likelihood estimates and posterior mode estimates for the unknown parameters of the modified network model using training data.
6. The method of claim 1 , further comprising:
applying an expectation-maximization (EM) algorithm to iterate execution of the modified network model using a content of the plurality of event-and-time specific texts;
wherein the EM algorithm computes an expectation of a log-likelihood for unknown parameters; and
wherein the unknown parameters comprise a plurality of states.
7. The method of claim 1 , further comprising:
dividing the plurality of event-and-time-specific texts into segments based on a state associated with a particular event-and-time-specific text and a time associated with a particular event-and-time-specific text; and
wherein the dividing of the plurality of event-and-time-specific texts into segments is performed by applying a Viterbi algorithm.
8. The method of claim 1 , wherein determining the particular set of texts, from the plurality of event-and-time specific texts, comprises eliminating spam texts, texts that contain less than a minimum threshold of words, texts that contain more than a maximum threshold of words, and texts that do not meet a word repetition requirement.
9. The method of claim 1 ,
wherein the plurality of event-and-time-specific texts comprises texts provided by social media search or tracking sites, including social media services;
wherein the plurality of event-and-time-specific texts comprises texts generated by users; and
wherein the plurality of event-and-time-specific texts relate to one or more sport events, concerts, political events, social events, or celebrity personalities.
10. The method comprising:
receiving training data that comprise a plurality of event-and-time-specific texts that are contextually related to a plurality of events;
determining one or more sets of vocabulary from the training data;
based on the one or more sets of vocabulary and the training data, determining initial values for one or more parameters for a modified network model that defines a plurality of states for the plurality of events;
initializing the modified network model using the initial values for the one or more parameters for the modified network model;
iteratively processing the training data to refine the modified network model to identify one or more states for the modified network model;
wherein the method is performed by one or more computing devices.
11. One or more non-transitory computer-readable storage media storing instructions which, when executed by one or more computing devices, cause:
receiving training data that comprise a plurality of event-and-time-specific texts that are contextually related to a plurality of events;
identifying, by iteratively processing the training data, a plurality of different states, each state corresponding to a different event from the plurality of events;
generating a modified network model that comprises the plurality of states;
receiving additional data that comprise a plurality of additional event-and-time-specific texts that are contextually related to a particular event;
processing the additional data by applying the modified network model to the additional data to identify, within the plurality of additional event-and-time specific texts, a particular set of texts that belong to a particular state of the plurality of states; and
identifying, within the particular set of texts, one or more texts that are most representative of all texts in the particular set of texts that belong to the particular state.
12. The one or more non-transitory computer-readable storage media of claim 11 , storing additional instructions which, when executed by the one or more computing devices, cause:
generating a summary of an event that corresponds to the particular state, wherein the summary is based on the one or more texts that are most representative of all texts in the particular set of texts that belong to the particular state.
13. The one or more storage media of claim 11 , wherein the modified network model is any one of: a modified Hidden Markov Model (HMM) or a modified Dynamic Bayesian Networks Model (DBNM).
14. The one or more non-transitory computer-readable storage media of claim 11 , storing additional instructions which, when executed by the one or more computing devices, cause:
initializing one or more elements of a network model;
wherein iterative processing the training data to generate the modified network model comprises determining one or more parameters of the network model;
identifying a plurality of state-and-event-specific words for the event in a content of the plurality of event-and-time-specific texts; and
including the plurality of state-and-event-specific words in determining the plurality of states within the event and iterating executing of the modified network model using the content of the plurality of event-and-time specific texts.
15. The one or more non-transitory computer-readable storage media of claim 11 , storing additional instructions which, when executed by the one or more computing devices, cause
determining unknown parameters of the modified network model using a Baum-Welch algorithm; and
wherein the Baum-Welch algorithm computes maximum likelihood estimates and posterior mode estimates for the unknown parameters of the modified network model using training data.
16. The one or more non-transitory computer-readable storage media of claim 11 , storing additional instructions which, when executed by the one or more computing devices, cause:
applying an expectation-maximization (EM) algorithm to iterate execution of the modified network model using a content of the plurality of event-and-time specific texts;
wherein the EM algorithm computes an expectation of a log-likelihood for unknown parameters; and
wherein the unknown parameters comprise a plurality of states.
17. The one or more non-transitory computer-readable storage media of claim 11 , storing additional instructions which, when executed by the one or more computing devices, cause:
dividing the plurality of event-and-time-specific texts into segments based on a state associated with a particular event-and-time-specific text and a time associated with a particular event-and-time-specific text; and
wherein the dividing of the plurality of event-and-time-specific texts into segments is performed by applying a Viterbi algorithm.
18. The one or more non-transitory computer-readable storage media of claim 11 , storing additional instructions which, when executed by the one or more computing devices, cause: eliminating spam texts, texts that contain less than a minimum threshold of words, texts that contain more than a maximum threshold of words, and texts that do not meet a word repetition requirement.
19. The one or more non-transitory computer-readable storage media of claim 11 :
wherein the plurality of event-and-time-specific texts comprises texts provided by social media search or tracking sites, including social media services;
wherein the plurality of event-and-time-specific texts comprises texts generated by users; and
wherein the plurality of event-and-time-specific texts relate to one or more sport events, concerts, political events, social events, or celebrity personalities.
20. A text processing system comprising at least one content server, wherein the at least one content server is configured to perform:
receiving training data that comprise a plurality of event-and-time-specific texts that are contextually related to a plurality of events;
identifying, by iteratively processing the training data, a plurality of different states, each state corresponding to a different event from the plurality of events;
generating a modified network model that comprises the plurality of states;
receiving additional data that comprise a plurality of additional event-and-time-specific texts that are contextually related to a particular event;
processing the additional data by applying the modified network model to the additional data to identify, within the plurality of additional event-and-time specific texts, a particular set of texts that belong to a particular state of the plurality of states; and
identifying, within the particular set of texts, one or more texts that are most representative of all texts in the particular set of texts that belong to the particular state.
21. The text processing system of claim 20 , wherein the at least one content server is further configured to perform:
generating a summary of an event that corresponds to the particular state, wherein the summary is based on the one or more texts that are most representative of all texts in the particular set of texts that belong to the particular state.
22. The text processing system of claim 20 , wherein the modified network model is any one of: a modified Hidden Markov Model (HMM) or a modified Dynamic Bayesian Networks Model (DBNM).
23. The text processing system of claim 20 , wherein the at least one content server is further configured to perform:
initializing one or more elements of a network model;
wherein iteratively processing the training data to generate the modified network model comprises determining one or more parameters of the network model;
identifying a plurality of state-and-event-specific words for the event in a content of the plurality of event-and-time-specific texts; and
including the plurality of state-and-event-specific words in determining the plurality of states within the event and iterating executing of the modified network model using the content of the plurality of event-and-time specific texts.
24. The text processing system of claim 20 , wherein the at least one content server is further configured to perform:
determining unknown parameters of the modified network model using a Baum-Welch algorithm; and
wherein the Baum-Welch algorithm computes maximum likelihood estimates and posterior mode estimates for the unknown parameters of the modified network model using training data.
25. The text processing system of claim 20 , wherein the at least one content server is further configured to perform:
applying an expectation-maximization (EM) algorithm to iterate execution of the modified network model using a content of the plurality of event-and-time specific texts;
wherein the EM algorithm computes an expectation of a log-likelihood for unknown parameters; and
wherein the unknown parameters comprise a plurality of states.
26. The text processing system of claim 20 , wherein the at least one content server is further configured to perform:
dividing the plurality of event-and-time-specific texts into segments based on a state associated with a particular event-and-time-specific text and a time associated with a particular event-and-time-specific text; and
wherein the dividing of the plurality of event-and-time-specific texts into segments is performed by applying a Viterbi algorithm.
27. The text processing system of claim 20 , wherein the at least one content server is further configured to perform:
eliminating spam texts, texts that contain less than a minimum threshold of words, texts that contain more than a maximum threshold of words, and texts that do not meet a word repetition requirement.
28. The text processing system of claim 20 ,
wherein the plurality of event-and-time-specific texts comprises texts provided by social media search or tracking sites, including social media services;
wherein the plurality of event-and-time-specific texts comprises texts generated by users; and
wherein the plurality of event-and-time-specific texts relate to one or more sport events, concerts, political events, social events, or celebrity personalities.