NavyTime: Event and Time Ordering from Raw Text Nate Chambers, US Naval Academy
Challenge Given a document of raw unlabeled text, label all of the words that represent events and time expressions. Then identify and label temporal relations between pairs of events and times.
NavyTime Overview A complete system with the following independent processes: 1. Event extraction 2. Event attribute extraction 3. Timex extraction with SUTime 4. Identifies pairs of events/times likely to have a temporal relation 5. Labels the actual temporal relation
4 Step Example 1. He drove home and ate dinner yesterday. (timex, 06/14/2013)
2. He drove home and ate dinner yesterday. (ev1, past, occur)
(ev2, past, occur)
3. He drove home and ate dinner yesterday.
4. He drove home and ate dinner yesterday. BEFORE
2. Event Extraction
1. Time Extraction We used Stanfordâ€™s SUTime system for token span identification (top system).
The #2 extraction system on 3 of the 4 aspects of event extraction. Takeaway: The same features were used for all 4 subtasks. MaxEnt classifier.
We created new rules for fiscal quarter reasoning to enhance SUTimeâ€™s performance. Ours is the second best system for time normalization.
3-4. Temporal Relation Extraction Task ABC: the #2 best performing system. We used several classifiers for specific tasks.
3. Relation Identification
Task ABC Results
Determine which pairs should have a relation. We evaluated rule-based and data-driven classifiers. The best setup: 1. event-event: MaxEnt 2. Event-DCT: MaxEnt 3. Event-time: rule-based
Raw Text -> Relations
4. Relation Labeling NavyTime separates relation decisions into independent classifiers. 1. event-DCT 2. Event-event same sentence 3. Event-event diff sentence 4. Event-timex same sentence 5. Event-timex diff sentence
1. 2. 3. 4.
Cleartk NavyTime JU-CSE KUL
30.98 27.28 24.61 19.01
Coming soon! Publicly released software, in a completely new architecture, planned for this Fall.