fine grained temporal relation extraction
play

Fine-Grained Temporal Relation Extraction Siddharth Vashishtha - PowerPoint PPT Presentation

Fine-Grained Temporal Relation Extraction Siddharth Vashishtha Benjamin Van Durme Aaron Steven White University of Rochester Johns Hopkins University University of Rochester Data and code available at: http://decomp.io Overarching


  1. Background Methodology Model Results Analysis Conclusion Representing Event Timelines - A novel Universal Decompositional Semantics (UDS) framework for temporal relation representation that puts event duration front and center. - We map the events or situations to a timeline represented in real numbers. Sam broke the window and ran away. broke 5 20 ran 25 60 0 reference-interval 100

  2. Background Methodology Model Results Analysis Conclusion Protocol Design - We ask questions about the chronology of events and the duration of each event - Annotated example (next slide)

  3. start-point end-point

  4. Background Methodology Model Results Analysis Conclusion Data Collection - We took English Web Treebank (EWT) from Universal Dependencies (UD) and designed a protocol to extract fine-grained temporal relations.

  5. Background Methodology Model Results Analysis Conclusion Data Collection - We took English Web Treebank (EWT) from Universal Dependencies (UD) and designed a protocol to extract fine-grained temporal relations. - Extracted predicates from UD-data using PredPatt (White et al., 2016; Zhang et al., 2017)

  6. Background Methodology Model Results Analysis Conclusion Constructed Data - We recruited 765 annotators from Amazon Mechanical Turk to annotate predicate pairs in groups of five. The resulting dataset is UDS-Time.

  7. Background Methodology Model Results Analysis Conclusion Constructed Data - We recruited 765 annotators from Amazon Mechanical Turk to annotate predicate pairs in groups of five. The resulting dataset is UDS-Time.

  8. Background Methodology Model Results Analysis Conclusion Constructed Data - We recruited 765 annotators from Amazon Mechanical Turk to annotate predicate pairs in groups of five. The resulting dataset is UDS-Time. 70k ~30k

  9. Background Methodology Model Results Analysis Conclusion Data Distributions Event Durations

  10. Background Methodology Model Results Analysis Conclusion Data Distributions Event Durations

  11. Background Methodology Model Results Analysis Conclusion Data Distributions Event Durations

  12. Background Methodology Model Results Analysis Conclusion Data Distributions Event Relations

  13. Background Methodology Model Results Analysis Conclusion Data Distributions Event Relations High Priority: High Priority: e1 Try googling it or type it into Try googling it or type it into youtube you might get lucky. youtube you might get lucky. e2

  14. Background Methodology Model Results Analysis Conclusion Data Distributions Event Relations High Containment: Both Tina and Vicky are excellent. I will definitely refer my friends and family. e2 e1

  15. Background Methodology Model Results Analysis Conclusion Data Distributions Event Relations High Equality: e1 I go Disco dancing and Cheerleading. It's fab! e2

  16. Background Methodology Model Results Analysis Conclusion Data Distributions Event Relations

  17. Background Methodology Model Results Analysis Conclusion Model

  18. Background Methodology Model Results Analysis Conclusion Goal To model the pairwise fine-grained temporal relations and durations by attempting to automatically build featural representations of each predicate, its duration and its relation.

  19. Background Methodology Model Results Analysis Conclusion Model Architecture 1. Event representation 2. Duration representation 3. Relation representation

  20. Background Methodology Model Results Analysis Conclusion Model Architecture 1. Event representation What to feed my dog after gastroenteritis? My dog has been sick for about 3 days now.

  21. Background Methodology Model Results Analysis Conclusion Model Architecture 1. Event representation What to feed my dog after gastroenteritis? My dog has been sick for about 3 days now.

  22. Background Methodology Model Results Analysis Conclusion Model Architecture 2. Duration representation What to feed my dog after gastroenteritis? My dog has been sick for about 3 days now.

  23. Background Methodology Model Results Analysis Conclusion Model Architecture 2. Duration representation What to feed my dog after gastroenteritis? My dog has been sick for about 3 days now.

  24. Background Methodology Model Results Analysis Conclusion Model Architecture 3. Relation representation What to feed my dog after gastroenteritis? My dog has been sick for about 3 days now.

  25. Background Methodology Model Results Analysis Conclusion Model Architecture 3. Relation representation What to feed my dog after gastroenteritis? My dog has been sick for about 3 days now.

  26. Background Methodology Model Results Analysis Conclusion Model Architecture Full Architecture What to feed my dog after gastroenteritis? My dog has been sick for about 3 days now.

  27. Background Methodology Model Results Analysis Conclusion Results

  28. Background Methodology Model Results Analysis Conclusion Performance on UDS-Time (test set) - We test 6 different variants of our model on the test set of UDS-Time

  29. Background Methodology Model Results Analysis Conclusion Performance on UDS-Time (test set) - We test 6 different variants of our model on the test set of UDS-Time

  30. Background Methodology Model Results Analysis Conclusion Performance on UDS-Time (test set) - We test 6 different variants of our model on the test set of UDS-Time

  31. Background Methodology Model Results Analysis Conclusion Performance on TimeBank-Dense A transfer learning approach on TimeBank-Dense to predict standard categorical temporal relations

  32. Background Methodology Model Results Analysis Conclusion Performance on TimeBank-Dense A transfer learning approach on TimeBank-Dense to predict standard categorical temporal relations. Features

  33. Background Methodology Model Results Analysis Conclusion Performance on TimeBank-Dense A transfer learning approach on TimeBank-Dense to predict standard categorical temporal relations.

  34. Background Methodology Model Results Analysis Conclusion Performance on TimeBank-Dense A transfer learning approach on TimeBank-Dense to predict standard categorical temporal relations. 0.566 0.529 0.519 0.494

  35. Background Methodology Model Results Analysis Conclusion Performance on TimeBank-Dense A transfer learning approach on TimeBank-Dense to predict standard categorical temporal relations. 0.566 0.529 0.519 0.494 Our transfer learning approach beats most systems on TimeBank-Dense ( Event-Event Relations)

  36. Background Methodology Model Results Analysis Conclusion Document Timelines - A model to induce document timelines from the pairwise predictions

  37. Background Methodology Model Results Analysis Conclusion Document Timelines - A model to induce document timelines from the pairwise predictions - The Spearman correlation for timelines induced from our model and the timelines induced from the actual data: beginning point : 0.28 duration : -0.097

  38. Background Methodology Model Results Analysis Conclusion Document Timelines - A model to induce document timelines from the pairwise predictions - The Spearman correlation for timelines induced from our model and the timelines induced from the actual data: beginning point : 0.28 duration : -0.097 - The low correlation values suggest that even though the model is good at predicting pairwise predictions, it struggles to generate the entire document timeline

  39. Background Methodology Model Results Analysis Conclusion Model Analysis

  40. Background Methodology Model Results Analysis Conclusion Which words are attended to the most? - We looked at the top 15 words in UDS-Time development set which have the highest mean duration-attention and relation-attention weights.

  41. Background Methodology Model Results Analysis Conclusion Which words are attended to the most? - Duration - We looked at the top 15 words in UDS-Time development set which have the highest mean duration-attention and relation-attention weights.

  42. Background Methodology Model Results Analysis Conclusion Which words are attended to the most? - Duration - We looked at the top 15 words in UDS-Time development set which have the highest mean duration-attention and relation-attention weights. - Words that denote some time period (months, minutes, hour etc.) have the highest mean duration attention-weights.

  43. Background Methodology Model Results Analysis Conclusion Which words are attended to the most? - Relation - We looked at the top 15 words in UDS-Time development set which have the highest mean duration-attention and relation-attention weights.

  44. Background Methodology Model Results Analysis Conclusion Which words are attended to the most? - Relation - We looked at the top 15 words in UDS-Time development set which have the highest mean duration-attention and relation-attention weights. - Words that are either coordinators (such as or and and ) , or bearers of tense information - i.e. lexical verbs and auxiliaries, have the highest mean relation attention weights

  45. Background Methodology Model Results Analysis Conclusion Which words are attended to the most? - Relation - We looked at the top 15 words in UDS-Time development set which have the highest mean duration-attention and relation-attention weights. - Words that are either coordinators (such as or and and ) , or bearers of tense information - i.e. lexical verbs and auxiliaries, have the highest mean relation attention weights

  46. Background Methodology Model Results Analysis Conclusion Conclusion

  47. Background Methodology Model Results Analysis Conclusion Introduction - Overarching question: How do humans extract chronology of events?

  48. Background Methodology Model Results Analysis Conclusion Introduction - Overarching question: How do humans extract chronology of events? Background - A standard approach in previous corpora: Categorical temporal relations

  49. Background Methodology Model Results Analysis Conclusion Introduction - Overarching question: How do humans extract chronology of events? Background - A standard approach in previous corpora: Categorical temporal relations - Limitations: no duration information, hard to annotate, lacking fine-grained relation distinctions

  50. Background Methodology Model Results Analysis Conclusion Introduction - Overarching question: How do humans extract chronology of events? Background - A standard approach in previous corpora: Categorical temporal relations - Limitations: no duration information, hard to annotate, lacking fine-grained relation distinctions Methodology: A new approach

  51. Background Methodology Model Results Analysis Conclusion Introduction - Overarching question: How do humans extract chronology of events? Background - A standard approach in previous corpora: Categorical temporal relations - Limitations: no duration information, hard to annotate, lacking fine-grained relation distinctions Methodology: A new approach - Mapping events to timelines represented in real number

  52. Background Methodology Model Results Analysis Conclusion Introduction - Overarching question: How do humans extract chronology of events? Background - A standard approach in previous corpora: Categorical temporal relations - Limitations: no duration information, hard to annotate, lacking fine-grained relation distinctions Methodology: A new approach - Mapping events to timelines represented in real number - Explicitly annotating event durations

  53. Background Methodology Model Results Analysis Conclusion Introduction - Overarching question: How do humans extract chronology of events? Background - A standard approach in previous corpora: Categorical temporal relations - Limitations: no duration information, hard to annotate, lacking fine-grained relation distinctions Methodology: A new approach - Mapping events to timelines represented in real number - Explicitly annotating event durations - Construction of a new dataset: UDS-Time

  54. Background Methodology Model Results Analysis Conclusion Model

  55. Background Methodology Model Results Analysis Conclusion Model - Vector representation of events, event-duration, fine-grained temporal relations

Recommend


More recommend