txtcat-nb is a simple text file classifier using naive bayes. A somewhat rudimentary description of the math involved is in techdoc.tex (but better descriptions of the technique exist elsewhere).

Training data is assumed to be in a single directory; subdirectories within the data directory used for labels.

For example,

ls data/ data/a data/b ls data/* data/a/1.txt data/a/2.txt data/b/3.txt data/b/4.txt

would correspond to a two class dataset with labels a and b, each with two data files.

python >= 2.7 (for collections.Counter and json). Also works with python3 (tested with 3.1.2).

For training data as above, storing trained classifier in model.json, using 30% of training files for each label for statistics:

train.py -d data/ -o model.json -p 0.3

or

train.py --datadir=data/ --output=model.json --pct=0.3

classify.py -m model.json -d unlabled-data/

or

classify.py --model=model.json --datadir=unlabled-data/

The smoothing constant can be changed if desired; the default of 1 seems to work reasonably well.

1. Doesn't currently support full cross-validation in training.
2. No pre-processing of words yet - accuracy for my current test cases is sufficient with no preprocessing.