Rule-based Morphological Inflection Improves Neural Terminology Translation

Data and source code for the EMNLP (2021) paper:

Rule-based Morphological Inflection Improves Neural Terminology Translation

Weijia Xu, Marine Carpuat

Test Suites

We construct two test suites en-de.health.tagged and en-lt.news.tagged to evaluate MT with lemma terminology constraints in domain adaptation and low-resource settings, respectively.

Code

We release the code for rule-based inflection in code/rule_based_inflection.py. It provides an inflect function that takes the following arguments:

• tgt_lang: language code of the target language (e.g. de for German and lt for Lithuanian)
• src_sentence: source sentence in a string
• src_terms: a list of source terms corresponding to each target lemma
• tgt_lemmas: a list of target lemmas to be inflected (should have the same length as src_terms)
• lemma_tag_to_form: a dictionary of dictionaries that maps a target lemma given a morphological tag to the inflected form. As an example, it may like this:

{
	"run": {
		"V,1per,plu,pres,ind": run,
		"V,1per,plu,past,ind": ran,
	},
	"point": {
		"NN,noGender,nom,sing": "point",
		"NN,noGender,nom,plu": "points"
	}
}

For German, we extract the dictionary from the German Morphological Dictionaries. For Lithuanian, we use UniMorph.