This document summarizes an approach for extracting biological events from text using a graph-based model. It describes using support vector machines (SVMs) to classify triggers of events and their arguments by analyzing features of tokens and their dependencies. Rich features are used, including capitalization, punctuation, stems, n-grams, and syntactic dependencies. The system achieves good but not state-of-the-art performance on the BioNLP 2009 shared task data for event extraction. Room for improvement is noted in using sequential models like CRFs rather than the strong independence assumptions of SVMs.

1. Extracting Complex Biological Events
with Rich Graph­Based Feature Sets
Jari Björne, Juho Heimonen, Filip Ginter, Antti
Airola, Tapio Pahikkala, Tapio Salakoski
BioNLP 2009 Workshop
Farzaneh Sarafraz
18 June 2009

2. BioNLP'09 Task 1
 Events in abstracts
 Given: gene and gene products (proteins)
 Wanted: events
− type
− trigger
− participant(s)
− cause (if applicable)

3. Example
"I kappa B/MAD­3 masks the nuclear localization
signal of NF­kappa B p65 and requires the
transactivation domain to inhibit NF­kappa B
p65 DNA binding. "
Event: negative regulation
Trigger: masks
Theme1: the first p65
Cause: MAD­3

4. Event Types
 Gene expression  Binding
 Transcription  Regulation
 Protein Catabolism  Positive regulation
 Localisation  Negative regulation
 Phosphorylation

5. Training and Test Data
 Training data: 800 abstracts
 Development data: 150 abstracts
 Test data: 260 abstracts

6. The System
 Trigger recognition
− Methods similar to NER
− Classification
 Argument detection
− Graph edge selection
− Classification
 Semantic post­processing
− Rule­based

7. Trigger Detection
 Token labelling (one for each type and one ­)
 92% of triggers are single token
− Adjacent tokens form a trigger if they appear in the
training data
 Triggers that share a token:
− Combined class: gene expression/pos regulation
 A graph node for each trigger
− Not duplicated just yet

8. Classification ­ SVM
 Token features
− Binary: capitalisation, presence of punctuation or
numeric characters
− Stem
− Character bigrams and trigrams
− Token is known triggers in training data
− All the above for linear and dependency
“neighbours”

9. Classification ­ SVM
 Frequency features
− # of named entities
 In sentence
 In a linear window around the token
 Bag­of­words count of token texts in the sentence (?)
 Dependency chains
− Up to depth of 3 from the token are constructed
− At each depth both token and frequency features
− Plus dep type and sequence of dep types in chain

10. Two SVMs
 “Somewhat” different feature sets
 Combined weighted results
“This design should be considered an artifact of
the time­constrained, experiment­driven
development of the system rather than a
principled design”

11. Precision/Recall trade­off
 Undetected trigger ­­> undetected event
 All triggers have events in the training data ­­>
bias towards reporting an event for all detected
triggers
 Adjust P/R explicitly
− multiply the negative class by β
− find β experimentally

12. Edge Detection
 Multi­class SVM
 All potential directed edges
− Event node to named entity
− Event node to event node (nested event)
− Labelled as theme, cause, or negative
 Each edge is predicted independently

13. Feature Set – Central Concept
Shortest undirected
path of syntactic
dependencies in the
Stanford scheme
parse of the
sentence.

14. Feature Set
 Token text, POS, entity/event class,
dependency (subject)
 N­grams: merging the attributes of 2­4
− Consecutive tokens
− Consecutive dependencies
− Each token and two neighbouring dependencies
− Each dependency and two neighbouring tokens
− One bigram showing direction

15. Other Features
 Individual component features
 Semantic node features
 Frequency features

16. Semantic Post­Processing
 Duplicate nodes
− Same class and same trigger
− Combined trigger
 Remove improper arguments
 Remove directed cycles by removing the
weakest link

17. Duplicating Event Nodes
 Task restrictions
− Two causes,
− must have theme,
− etc.
 Several heuristics
 x­th first dependency
in shortest path from
the event for binding

18. Results

19. Compared to Us

20. What Didn't Work/Wasn't Tried
 CRF
 HMM
 Removing strong independence assumption
 Co­reference resolution (4.8%)

21. End.