Zachary May
CPH 738
Naive Bayes Text Classification
Files:
------
README This file.
classifiers.py Contains class definitions for standard and binary classifiers
as well as a testing harness for running tests on a trained
classifier and computing stastitics for those test runs.
nb.py The main program file.
test.py Unit tests. (These can be ignored.)
corpora/ Folder containing testing, training, and stop-word data sets.
Usage:
------
The program can be invoked with:
python nb.py
This runs the default scenario: training set 1, no stop word filtering, using
the non-binary version of the algorithm.
The command-line arguments the script takes differ only slightly from the
description in the assignment to work better with Python's standard argument
parsing library. Run the program with the --help switch for specifics.
Results:
--------
The following table is an overview of the statistics for both normal and binary
classifiers, trained on both training sets, with or without stop-word
filtering. I was somewhat surprised at how poorly the binary classifier
performed when trained with the larger corpus.
Training Set | Features | Prec | Recl | F-Score
---------------|-------------------------------|------|------|---------
training1.csv | unigrams | 0.74 | 0.95 | 0.83
training1.csv | unigrams, excl. stop words | 0.75 | 0.96 | 0.84
training1.csv | unigrams, binary | 0.76 | 0.94 | 0.84
training1.csv | unigrams, binary, excl. words | 0.74 | 0.94 | 0.83
training2.csv | unigrams | 0.90 | 0.91 | 0.90
training2.csv | unigrams, excl. stop words | 0.90 | 0.91 | 0.90
training2.csv | unigrams, binary | 0.86 | 0.86 | 0.86
training2.csv | unigrams, binary, excl. words | 0.88 | 0.86 | 0.87