Lemma Split Data for Morphological Inflection

Repo for the paper (Un)solving Morphological Inflection: Lemma Overlap Artificially Inflates Models' Performance.

Contents

• LemmaSplitData contains a lemma split version of the data from SIGMORPHON's 2020 shared task 0.
• lstm contains the baseline LSTM model.
• generate_lemma_splits.py is the script used to produce the data in practice.

Generating Lemma-Splits Data

1. Clone the SIGMORPHON 2020 task 0 data - the 3 folders DEVELOPMENT_LANGUAGES, SURPRISE_LANGUAGES and GOLD-TEST - to the folder DataExperiments/FormSplit.
2. Run the script generate_lemma_splits.py. It will generate a folder called DataExperiments/LemmaSplit at the same level and the same families sub-division (without the covered test files), where the samples are split across lemmas instead of randomly.