JavaNLP meeting notes for 05/07/03

Today we talked about general prospects for the code currently in JavaNLP
and Dan gave a tutorial on the new classify API. We meet again next week
(5/14) at our usual time (4-5pm, Gates 498).

JavaNLP stats (excluding pubcrawl):
- 806 Java files in edu.stanford.nlp 
- 173374 lines of Java code total
- Package with most classes: trees (96)
- Package with fewest classes: ie.ner (1)
- Package with most lines of code: redwoods (26135)
- Package with fewest lines of code: ie.ui (218)
- Longest Java file: lexparser.lexparser.FactoredParser (7800)
- Shortest Java file: util.XScored (6)

Notes on various packages:
- cluster: rewrite (Sep's oldest code / broken)
- database: just use Maps?
- fsm: move with oldparser
- lexgram: move with oldparser
- linalg: data structures to abstract over matrix package (e.g. colt)
- mt (does alignment, need decoder for full mt system)
- tagger: need generic tagging API
- wsd: "mess" of code for senseval, no real WSD code

Highlights of classify tutorial:
- code is in classify package and uses Datum/DataCollection from dbm
- use ClassifierFactory's trainClassifier method to make a Classifier
- use Classifier's classOf(Datum) to perform classification
- use Classifier's scoresOf(Datum) to get a Counter from classes to scores
- subclass AbstractLinearClassifierFactory if you can for new classifiers

Classify todo:
- Dan: make Classifier and ClassifierFactory serializable
- Dan: prior for Naïve Bayes classifier
- Dan/Joseph: use DataCollection code in classify code where appropriate
- Kristina: implement DecisionTree with splitting on feature present/absent
- Kristina: implement Perceptron over fixed feature vocab (1/0 or freq)
- Joseph/Dan: ClassifierTester package for acc/prec/recall/etc

See you next week! 
Thanks, js