Unsupervised Cross-lingual Matching of Product Classifications

Abstract

Unsupervised cross-lingual embeddings mapping has provided a unique tool for completely unsupervised trans- lation even for languages with different scripts. In this work we use this method for the task of unsupervised cross-lingual matching of product classifications. Our work also investigates limitations of unsupervised vector alignment and we also suggest several techniques for aligning product classifications based on their descriptions: using or discarding hierarchical information, using translations or pre-matched cross-lingual embeddings and all possible combinations of these approaches.

Keywords

• cross-lingual embeddings
• word embeddings
• unsupervised category matching