TODO.txt

- Tune prec, recall (boost precision)
- Feature: was the word seen before
- Feature: length for date entities, and for named entities
- Better cost?
- Debug dev reporting
- Try JAMR old weights on new data
- Ask about ISI result
- Retrain relation ID

Tues May 26
- Add regularizer
- Train relation id w/ automatic concepts

Wed May 27
- Correct normalization when checking spans (will matter for dates especially)
- Cost func shouldn't care about entity type
- Boost recall / play with reg
- Cost function should be this: if you can predict the correct concept, the predict it, otherwise predicting anything is better than nothing (this cost func depends on the output set!)

- WordNet pass through and POS
- log(Count) of the word (or log(Pr(produce something)))
- Add lists for NE, and unigrams for NE, does it start with capital
- WordNet pass through and log(count) of the word
- Syntactic context with concept (for sense)
- Is the concept a -NN (and w/ POS)
- run SRL to help event concept id
- parsed OntoNotes examples? look at possible roles and syntactic context 


Wed June 10
- Fix ordering of NEs
- Train stage2 w/ output of stage1 (w/ folds? or without)

Thurs June 11
- Add regularizer
- Try regular ramp (compare to infinite ramp)