Unsupervised Cross-Lingual Part-of-Speech Tagging with Monolingual Corpora Only

Researchers have developed a fully unsupervised framework for cross-lingual part-of-speech tagging that requires only monolingual text, eliminating the need for scarce parallel corpora. The method uses unsupervised neural machine translation to create pseudo-parallel sentences, then projects POS tags via word alignment. Tested on 28 language pairs, it matches or exceeds baseline performance that uses parallel data. A novel multi-source projection technique provides an average 1.3% performance improvement.