Code and data for the paper "Evaluating Morphological Generalisation in Machine Translation by Distribution-Based Compositionality Assessment"

Note: for updates, see https://github.com/aalto-speech/dbca

Structure

• scripts numbered 01-13 are meant to be run in succession
• run.sh provides examples of running the scripts
• exp/subset-d-1m/data contains the 1M sentence pair dataset
• exp/subset-d-1m/splits/*/*/*/ids_{train,test_full}.txt.gz contain the data splits with different compound divergences and different random initialisations

Dependencies

• Data is from the Tatoeba Challenge data release (eng-fin set)
• Data filtering is done using OpusFilter
• Morphological parsing is done using TNPP, CoNLL-U format parsed using this parser
• Data split algorithm uses PyTorch
• Tokenisers are trained using sentencepiece
• Translation systems are trained with OpenNMT-py
• Evaluating translations is done with sacreBLEU