E. Buyko, E. Faessler, J. Wermter, and U. Hahn. Proceedings of the BioNLP 2009 Workshop Companion Volume for Shared Task, page 19--27. Boulder, Colorado, Association for Computational Linguistics, (June 2009)
Description
BioNLP'09 event extraction by pruning dependency graph. Seems designed to handle most common cases well, ignoring others. 2nd best system in shared task eval.
Manually constructed dictionary of lemmatised trigger terms; most non-discriminative terms were rejected. Maximum-likelihood disambiguation for multiple triggers in same sentence.
Dependency graphs are pruned to (a) remove irrelevant lexical nodes (aux and modals); and (b) perform lexical-semantic generalisation (by lookup in various lexicons).
Argument id: variant approaches for easy event types, binding (ignoring > 2 themes) and regulation. Used a feature-based classifier and a graph kernel (with variants) in combination, as determined on development data.
Sentence-level only.
Label triggers in original training data in order to generate more negative examples for event determination.
Most errors are false arguments, or no trigger found (in particular, words that are double-annotated).
%0 Conference Paper
%1 buyko2009
%A Buyko, Ekaterina
%A Faessler, Erik
%A Wermter, Joachim
%A Hahn, Udo
%B Proceedings of the BioNLP 2009 Workshop Companion Volume for Shared Task
%C Boulder, Colorado
%D 2009
%I Association for Computational Linguistics
%K biomedical bionlp09 corpus_genia dependency_parse event_extraction
%P 19--27
%T Event Extraction from Trimmed Dependency Graphs
%U http://www.aclweb.org/anthology/W09-1403
@inproceedings{buyko2009,
added-at = {2009-10-22T14:10:11.000+0200},
address = {Boulder, Colorado},
author = {Buyko, Ekaterina and Faessler, Erik and Wermter, Joachim and Hahn, Udo},
biburl = {https://www.bibsonomy.org/bibtex/273bd9e1a295cae745cdb2c451cd2778a/jnothman},
booktitle = {Proceedings of the BioNLP 2009 Workshop Companion Volume for Shared Task},
description = {BioNLP'09 event extraction by pruning dependency graph. Seems designed to handle most common cases well, ignoring others. 2nd best system in shared task eval.
Manually constructed dictionary of lemmatised trigger terms; most non-discriminative terms were rejected. Maximum-likelihood disambiguation for multiple triggers in same sentence.
Dependency graphs are pruned to (a) remove irrelevant lexical nodes (aux and modals); and (b) perform lexical-semantic generalisation (by lookup in various lexicons).
Argument id: variant approaches for easy event types, binding (ignoring > 2 themes) and regulation. Used a feature-based classifier and a graph kernel (with variants) in combination, as determined on development data.
Sentence-level only.
Label triggers in original training data in order to generate more negative examples for event determination.
Most errors are false arguments, or no trigger found (in particular, words that are double-annotated).},
interhash = {00300389d6c4516031ddbe6e54a21360},
intrahash = {73bd9e1a295cae745cdb2c451cd2778a},
keywords = {biomedical bionlp09 corpus_genia dependency_parse event_extraction},
month = {June},
pages = {19--27},
publisher = {Association for Computational Linguistics},
timestamp = {2009-10-22T14:10:11.000+0200},
title = {Event Extraction from Trimmed Dependency Graphs},
url = {http://www.aclweb.org/anthology/W09-1403},
year = 2009
}