Port from Python2 to Python3.

[meyersbs]

Changes:

Python:

• General syntax updates for changes from Python2 to Python3
• Code comments have been significantly updated
• cPickle --> pickle

Dependencies:

• scipy==0.12.0 --> scipy==0.18.1
• scikit-learn==0.15.1 --> scikit-learn==0.18.1
• numpy==1.9.0 --> numpy==1.12.0
• nltk==3.0.0 --> nltk==3.2.1

Support Data:

• There was a since nonstandard unicode character (the accented i in naive) in the list of negative words. I replaced this character with the correct escape sequence. To properly parse the escape sequence, codecs.open was used instead of the standard open for this file.

Additions:

Classification Model:

• The use updated pickle and scikit-learn libraries caused quite a hassle:
  • Classification models created using scikit-learn prior to version 0.18.1, did not embed the scikit-learn version into model. Since 0.18.1 does embed its version number, it gave a number of warnings about loading an outdated model.
  • I tried countless methods, but Python3's pickle simply cannot derserialize an object that was serialized using Python2's cPickle.
• In response to these conjoined errors, I decided to retrain the classification model using the original dataset. Since this required generating dependency parses for each request in the dataset, I chose to dump the annotated dataset, with dependency parses, as a json dictionary in the format expected by /scripts/train_model.py. See the README for information about how to download these datasets (they are too large for version control).
• The retrained model is included in this pull-request. Since the json formatted training datasets I have provided are so large, retraining the model will require a significant amount of memory.

Original Datasets:

• I have included the original datasets in the corpora/ directory.

Helper Functions:

• I have created an additional file, corpora/download.py, which automates downloading and extracting the json formatted training data.