Universal language segment representations learning with conditional masked language model

Abstract

The present disclosure provides a novel sentence-level representation learning method Conditional Masked Language Modeling (CMLM) for training on large scale unlabeled corpora. CMLM outperforms the previous state-of-the-art English sentence embedding models, including those trained with (semi-)supervised signals. For multilingual representations learning, it is shown that co-training CMLM with bitext retrieval and cross-lingual natural language inference (NL) fine-tuning achieves state-of-the-art performance. It is also shown that multilingual representations have the same language bias and principal component removal (PCR) can eliminate the bias by separating language identity information from semantics.