This is the code for Pointer-Generator Networks for Low-Resource Machine Translation: Don't Copy That!, presented at LREC-COLING '24.
Train MT with PGNs like this:
python pgn.py \
--DATADIR_L1 /export/b08/nbafna1/data/nllb/hin-bho/splits/hin \
--DATADIR_L2 /export/b08/nbafna1/data/nllb/hin-bho/splits/bho \
--TOKENIZER_INPATH $TOKENIZER_INPATH \
--OUTPUT_DIR $MODEL_OUTPUT_DIR --LOG_DIR $LOG_DIR --epochs $epochs --batch_size $batch_size \
--max_lines $max_lines
$TOKENIZER_INPATH is the path to a HuggingFace AutoTokenizer, L1 and L2 are the source and target languages respectively.