Winning solution for the Rakuten Data Challenge, as part of SIGIR eCom '18.
The details of the model choices and evolution can be found in the system description paper for that workshop.
Set up the expected data directories, from the repository root:
mkdir -p data/models
Move the challenge files into the data/ subdirectory:
mv path/to/rdc-catalog-train.tsv data/
mv path/to/rdc-catalog-test.tsv data/
Run a train/test split, build the vocabularies, and save the int-encoded training and validation sets for later:
./prep.sh
Train and save a forward model with the hyperparameters from the winning RDC solution (the model goes in data/models/model-name.h5):
./train.sh model-name
Train a reverse model, intended for use in building a bi-directional ensemble with a forward network:
./train.sh reverse-model --reverse
Train a model with some parameters different from the default:
./train.sh custom-model-name --n-epochs=20 --lr=1.2
See a full list of parameters available to tune via flags:
./train.sh -- --help
Run an inference on the validation set, generate predictions, and then output precision, recall, and F1:
./infer.sh model-name
./infer.sh --forward=model-name # equivalent
Score a reverse model:
./infer.sh --reverse=reverse-model
Similarly for a bi-directional ensemble:
./infer.sh --forward=model-name --reverse=reverse-model
Or for a larger ensemble, e.g. with 4 each forward and reverse:
./infer.sh --forward=fwd1,fwd2,fwd3,fwd4 --reverse=rev1,rev2,rev3,rev4
Since that can take awhile, you can show intermediate results along the way:
./infer.sh --forward=fwd1,fwd2,fwd3,fwd4 --reverse=rev1,rev2,rev3,rev4 --debug
To run test set inference and output prediction files for a single model, with ensembles working analogously to the commands above:
./infer.sh model-name --is-test