Event Extraction as Dependency Parsing (in BioNLP 2011) David - - PowerPoint PPT Presentation

▶
event extraction as dependency parsing in bionlp 2011
SMART_READER_LITE
LIVE PREVIEW

Event Extraction as Dependency Parsing (in BioNLP 2011) David - - PowerPoint PPT Presentation

Jul 09, 2023 166 likes •425 views

Event Extraction as Dependency Parsing (in BioNLP 2011) David McClosky Stanford University 6.24.2011 Joint work with Mihai Surdeanu and Christopher D. Manning Summary Event parsing Our approach in two slides... David McClosky (Stanford)

slide-1
SLIDE 1

Event Extraction as Dependency Parsing (in BioNLP 2011)

David McClosky

Stanford University

6.24.2011 Joint work with Mihai Surdeanu and Christopher D. Manning

slide-2
SLIDE 2

Summary Event parsing

Our approach in two slides...

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 1 / 10

slide-3
SLIDE 3

Summary Event parsing

Our approach in two slides...

Full details in [McClosky, Surdeanu, and Manning, ACL 2011]

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 1 / 10

slide-4
SLIDE 4

Road map You are here.

Outline

1

Event Parsing

2

Adapting to BioNLP 2011

3

Experiments

4

Conclusion

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 2 / 10

slide-5
SLIDE 5

Event Parsing Overview

Approach

Preprocessing: Segmentation, tokenization

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 3 / 10

slide-6
SLIDE 6

Event Parsing Overview

Approach

Preprocessing: Segmentation, tokenization, syntactic parsing Self-trained biomedical parser: [McClosky, 2010]

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 3 / 10

slide-7
SLIDE 7

Event Parsing Overview

Approach

Anchor classification: Token classification for event anchors (similar to [Bj¨

  • rne et al., BioNLP 2009])

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 3 / 10

slide-8
SLIDE 8

Event Parsing Overview

Approach

Event parsing: Parse anchors and proteins using reranking parser

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 3 / 10

slide-9
SLIDE 9

Event Parsing Motivation

Maximum-spanning tree based parsing

Why a dependency parser? Event structures are non-projective (non-planar)

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 4 / 10

slide-10
SLIDE 10

Event Parsing Motivation

Maximum-spanning tree based parsing

Why a dependency parser? Event structures are non-projective (non-planar) Why MSTParser? [McDonald et al., EMNLP 2005] Handles non-projective trees naturally

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 4 / 10

slide-11
SLIDE 11

Event Parsing Motivation

Maximum-spanning tree based parsing

Why a dependency parser? Event structures are non-projective (non-planar) Why MSTParser? [McDonald et al., EMNLP 2005] Handles non-projective trees naturally Easy to extend feature extractor

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 4 / 10

slide-12
SLIDE 12

Event Parsing Motivation

Maximum-spanning tree based parsing

Why a dependency parser? Event structures are non-projective (non-planar) Why MSTParser? [McDonald et al., EMNLP 2005] Handles non-projective trees naturally Easy to extend feature extractor Support for n-best parsing

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 4 / 10

slide-13
SLIDE 13

Adapting to BioNLP 2011 Overview

Adapting to BioNLP 2011

General improvements

Distributional similarity features in anchor detection

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 5 / 10

slide-14
SLIDE 14

Adapting to BioNLP 2011 Overview

Adapting to BioNLP 2011

General improvements

Distributional similarity features in anchor detection Improved head percolation rules for multiword anchors

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 5 / 10

slide-15
SLIDE 15

Adapting to BioNLP 2011 Overview

Adapting to BioNLP 2011

General improvements

Distributional similarity features in anchor detection Improved head percolation rules for multiword anchors Using lemmas (along with word forms) during event parsing

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 5 / 10

slide-16
SLIDE 16

Adapting to BioNLP 2011 Overview

Adapting to BioNLP 2011

General improvements

Distributional similarity features in anchor detection Improved head percolation rules for multiword anchors Using lemmas (along with word forms) during event parsing

Domain-specific customization

Update event type information (EPI, ID)

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 5 / 10

slide-17
SLIDE 17

Adapting to BioNLP 2011 Overview

Adapting to BioNLP 2011

General improvements

Distributional similarity features in anchor detection Improved head percolation rules for multiword anchors Using lemmas (along with word forms) during event parsing

Domain-specific customization

Update event type information (EPI, ID) Combine ID training data with GENIA (ID)

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 5 / 10

slide-18
SLIDE 18

Adapting to BioNLP 2011 Overview

Adapting to BioNLP 2011

General improvements

Distributional similarity features in anchor detection Improved head percolation rules for multiword anchors Using lemmas (along with word forms) during event parsing

Domain-specific customization

Update event type information (EPI, ID) Combine ID training data with GENIA (ID) Removing nested entities (ID)

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 5 / 10

slide-19
SLIDE 19

Experiments

Results on Genia development

Decoder(s) Parser Reranker 1P 49.0 49.4 2P 49.5 50.5 1N 49.9 50.2 2N 46.5 47.9 All — 50.7

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 6 / 10

slide-20
SLIDE 20

Experiments

Results on Epigenetics development

Decoder(s) Parser Reranker 1P 62.3 63.3 2P 62.2 63.3 1N 62.9 64.6 2N 60.8 63.8 All — 64.1 (note: issues with our internal evaluator implementation)

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 7 / 10

slide-21
SLIDE 21

Experiments

Domain adaptation for Infectious Diseases

Model Precision Recall f-score

ID

59.3 38.0 46.3

ID (×1) + GE

52.0 40.2 45.3

ID (×2) + GE

52.4 41.7 46.4

ID (×3) + GE

54.8 45.0 49.4

ID (×4) + GE

55.2 43.8 48.9

ID (×5) + GE

55.1 44.7 49.4

(parser only with 2N decoder)

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 8 / 10

slide-22
SLIDE 22

Experiments

Results on Infectious Diseases development

Decoder(s) Parser Reranker 1P 46.0 48.5 2P 47.8 49.8 1N 48.5 49.4 2N 49.4 48.8 All — 50.2

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 9 / 10

slide-23
SLIDE 23

Conclusion Talks this short probably don’t need subsections...

Summary

New approach to event extraction

Parsing can be used for event extraction Reranker further improves performance

Minimal changes to adapt to new BioNLP domains Component in the FAUST system (stay tuned!) Code coming soon! http://nlp.stanford.edu/software/eventparsing.shtml

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 10 / 10

slide-24
SLIDE 24

Conclusion Talks this short probably don’t need subsections...

Summary

New approach to event extraction

Parsing can be used for event extraction Reranker further improves performance

Minimal changes to adapt to new BioNLP domains Component in the FAUST system (stay tuned!) Code coming soon! http://nlp.stanford.edu/software/eventparsing.shtml

Questions?

David McClosky (Stanford) Event Parsing in BioNLP 2011 6.24.2011 10 / 10