Evita: A Robust Event Recognizer For QA Systems
Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing, page 700--707. Vancouver, British Columbia, Canada, Association for Computational Linguistics, (October 2005)Description
System to recognise TimeML events, not specific event types. Determines anchor and grammatical attributes.
Events defined as "those event-denoting expressions that participate in the narrative of a given document and which can be temporally ordered," be they dynamic or static; VB, NN or AD.
Combines linguistic, statistical, pattern-based and knowledge-based components.
Trigger detection (after head and polarity/modal markers are separated):
- verbs: some heads (e.g. copula) are rejected. (It seems that otherwise, an event is assumed? "Lexical lookup" isn't elaborated!)
- noun: head is event if it falls under one of 25 WordNet subtrees, with Bayesian WSD in conflict.
- adj: only those annotated in TimeBank, when appearing as predicative complements.
Attributes:
- tense, aspect, non-finite morphology: 140 rules mapping verbal phrases to these features;
- modality, polarity: basic patterns within chunk;
- semantic class (state, occurrence, reporting, intention, perception): by majority class for word from training.
Only look for within-chunk attributes, so chunks are first clustered: triggered pattern matching.
Overall F-measure 80.12%, comparable to inter-annotator agreement. Attribute accuracy 86-98%, but may just reflect major classes.
Part of TARSQI toolkit focusing on time relations in text.