A Markov Logic Approach to Bio-Molecular Event Extraction
Proceedings of the BioNLP 2009 Workshop Companion Volume for Shared Task, page 41--49. Boulder, Colorado, Association for Computational Linguistics, (June 2009)Description
Learn a joint probabilistic model (not pipeline) with Markov Logic (Statistical Relation Learning language). 3-4th in BioNLP; best results for task 2.
Represent events as relational structures over the tokens of sentence (close to SRL). Essentially First-Order Logic formulae with weights. Markov Logic does not yet handle cases where the number and identity of entities is unknown.
Map event annotations into MLNs: E = (L, C) where L is a set of token-token links, and C is a set of event triggers.
Four hidden predicates (like target class features): event(i), eventType(i, t), site(i), role(i, j, r)
Other predicates (features) regard entity annotations, dependency paths (from Charniak and CCG), dictionary lookups.
Provide implication logic for features to target hidden predicates (each gets assigned a weight in training), and global constraints. Many global constraints help validate; additional ones are in response to error analysis (e.g. token cannot be argument of multiple events).
Note that constraints on role assignment can improve event trigger detection in a joint model.
Links and resources
Tags
community
@jnothman's tags highlighted