Unsupervised Cross-Lingual Mapping for Phrase Embedding Spaces

Cross-lingual embedding has shown an effective way to learn cross-lingual representation in a joint embedding space. Recent work showed that cross-lingual phrase embedding is important to induce phrase table for unsupervised phrase-based machine translation. However, most of the cross-lingual representation from the literature are either limited to word level embedding or uses bilingual supervision for shared phrase embedding space. Therefore, in this paper, we explore the ways to map phrase embeddings of two languages into a common embedding space without supervision. Our model uses a three-step process: first we identify phrase in a sentence by using their mutual information, and combine component words of the phrase in the preprocessing stage; then we independently learn phrase embedding for each language based on their distributional properties, finally a fully unsupervised linear transformation method based on self-learning is used to map the phrase embeddings into a shared space. We extracted bilingual phrase translation as a gold standard to evaluate the result of the system. Besides its simplicity, the proposed method has shown a promising result for phrase embedding mapping.