First of all, thanks for sharing this code of the PAN15 competition.

Second, how should I prepare my own dataset and then train caravel to find the author of my anonymous text?

I have been trying to compile charmodel.so but I keep having the next error:

$ make charmodel.so
cd recur && git pull
Already up-to-date.
cd recur && make charmodel.so
make[1]: Entering directory `/home/versae/git/caravel/recur'
python setup.py build_ext --inplace
running build_ext
building 'charmodel' extension
gcc-4.9 -fno-strict-aliasing -DNDEBUG -g -fwrapv -O2 -Wall -Wstrict-prototypes -fPIC -D_GNU_SOURCE -DVECTOR -I. -I/usr/include/python2.7 -c py-recur-text.c -o build/temp.linux-x86_64-2.7/py-recur-text.o -march=native -ggdb -std=gnu11 -I. -Wall -O3 -ffast-math -fno-inline -DVECTOR -DPIC
py-recur-text.c: In function ‘Alphabet_encode_text’:
py-recur-text.c:128:52: error: ‘alphabet’ undeclared (first use in this function)
     u8 *encoded_text = rnn_char_alloc_encoded_text(alphabet,
                                                    ^
py-recur-text.c:128:52: note: each undeclared identifier is reported only once for each function it appears in
error: command 'gcc-4.9' failed with exit status 1
make[1]: *** [charmodel.so] Error 1
make[1]: Leaving directory `/home/versae/git/caravel/recur'
make: *** [recur/charmodel.so] Error 2

I changed to g++-4.9 in the Makefile and export CC and CXX in order for Python to find it, but still failing.

Thanks.

I'm trying to reproduce the PAN15 setup for the English corpora right now, but I've got a couple of questions which I can't answer from either your paper or the source code which are:

• How do I train a model once, save it and then use it for quick classification? pan-test or pan-ensemble don't do that apparently.
• Which method (pan-test or pan-ensemble) was used for creating the final results and how long did the calculation take?
• What exactly is the control corpus? From the code I would say it's some additional training mass (e.g. PAN14 corpora).
• Which corpora (PAN15, PAN14) were used for training, control and testing?
• Why does pan-test pass the input testing corpus (-i) as training corpus to train-net and why does get_net_and_corpus() in train-net add given test texts to the training texts?
• Is there a reason why pan-test uses a hedging value of 0.0 while pan-ensemble uses 0.03 for English texts?

It would be great if you could help me with this. Thanks!