MACE: Morphology Aware Cross-Lingual Embedding Using Contrastive Learning

The bilingual dictionary induction (BDI) task induced by cross-lingual word embedding usually performs poorly when the language pairs are structurally non-isomorphic, like English-Manipuri. The performance also degrades when one of the languages in the pair is morphologically rich. Earlier work shows that segmenting the morpheme (root) and suffix/prefix using a morphological analyzer slightly improves cross-lingual word embedding in the BDI task. In this work, we proposed a novel contrastive learning method that takes advantage of the rich morphological nature of a language without segmenting root words and suffixes/prefixes. The proposed contrastive learning pulls the source and target words together in a bilingual dictionary and brings the target word with another target word that shares the same root. From various experimental observations over four language pairs, it is evident that the proposed method outperforms the baseline methods in the BDI, Machine Translation (MT), and Cross-lingual Sentence Retrieval Task (CLSRT).