From: Christopher Potts Date: July 30, 2011 23:35:59 MDT To: Class list Subject: Computational Pragmatics: team feature-functions available for review Computational Pragmaticists! I hid out from the afternoon heat and implemented the models presented by Team Banana Wugs and Team Banana Slugs.  If you like, you can download them to check for bugs and areas for potential improvement: # Team Banana Slugs http://compprag.christopherpotts.net/code-data/pdtb_competition_team_banana_slugs.py # Team Banana Wugs http://compprag.christopherpotts.net/code-data/pdtb_competition_team_banana_wugs.py I'll be using a pre-defined train/test split for the competition. Both the train and test sets are balanced with respect to the semantic categories we're predicting, which should prevent a take-over of Expansion and in turn allow all your features to have their day in the sun. Here's the training set: http://compprag.christopherpotts.net/code-data/pdtb-competition-implicit-train-indices.pickle If you download that too, then running python pdtb_competition_team_banana_wugs.py or python pdtb_competition_team_banana_slugs.py will print out partial string representations of the training-set examples along with the model's view of them, so that you can see what the model is doing (and perhaps spot errors). Those of you who want to use the banana_slugs features should also download and unpack the attached archive of genre files. (I'm not sure whether I can distribute these from the website.) If you explore both, you'll very quickly see dramatic differences in how the two teams approached the problem. I updated pdtb_classifier.py with various assessment methods and facilities for working with pre-defined train/test splits: http://compprag.christopherpotts.net/code-data/pdtb_classifier.py I also updated pdtb.py a bit, so download that again as well: http://compprag.christopherpotts.net/code-data/pdtb.py Studying the following page would get you to the point where you could test your model yourself on random train/test splits, which would likely give you a major edge: http://compprag.christopherpotts.net/pdtb-predict-implicit.html I'll accept improvements to the models until midnight on Monday, but the improvements have to be stated as working code revisions. On Tuesday, I'll report the competition results and present prizes to the winning team! ---Chris