English Contrastive Learning Can Learn Universal Cross-lingual Sentence Embeddings
share
Universal cross-lingual sentence embeddings map semantically similar cross-lingual sentences into a shared embedding space. Aligning cross-lingual sentence embeddings usually requires supervised cross-lingual parallel sentences. In this work, we propose mSimCSE, which extends SimCSE to multilingual settings and reveal that contrastive learning on English data can surprisingly learn high-quality universal cross-lingual sentence embeddings without any parallel data. In unsupervised and weakly supervised settings, mSimCSE significantly improves previous sentence embedding methods on cross-lingual retrieval and multilingual STS tasks. The performance of unsupervised mSimCSE is comparable to fully supervised methods in retrieving low-resource languages and multilingual STS. The performance can be further enhanced when cross-lingual NLI data is available. Our code is publicly available at https://github.com/yaushian/mSimCSE.
READ FULL TEXT