Timexla is a temporal expression detector, extractor, and resolver.
It seeks to be:
- less pre-packaged than SUTime, Stanford NLP's temporal expression recognizer, which is great but comes rolled into the whole 100+ MB CoreNLP package.
- less rule-based than HeidelTime, which is state of the art, but operates on a lot of effective, but hard-coded rules.
Two improvements, which are novel in the field as far as I know, are to:
- Power (temporal) anaphora resolution by carrying along an anchor/center of temporal location, much as one does with pronoun anaphora.
- Adjusting priors (leaning forward or backward) depending on tense as well as aspect of verbs connected to
You'll need the TimeML corpus to run the code as-is. Fortunately, this is free, or something vaguely resembling free (despite the $0 total), because you have to petition the LDC for it.
- Christopher Brown · mailto:io@henrian.com
- Justin Cope · mailto:justin.cope@utexas.edu
- Thanks to Jason Baldridge for some of the smoothing algorithms used in the HMM (Hidden Markov Model) algorithm.