OmniAlign Unified Model for Word and Sentence Alignment

OmniAlign uses an encoder-only backbone to perform both word-level and sentence-level cross-lingual alignment within a single lightweight model. It derives word alignments from contextualized token similarity and document-level alignments via sentence embeddings and dynamic programming.