XLEnt

XLEnt consists of parallel entity mentions in 120 languages aligned with English. These entity pairs were constructed by performing named entity recognition (NER) and typing on English sentences from mined sentence pairs. These extracted English entity labels and types were projected to the non-English sentences through word alignment. Word alignment was performed by combining three alignment signals ((1) word co-occurence alignment with FastAlign (2) semantic alignment using LASER embeddings, and (3) phonetic alignment via transliteration) into a unified word-alignment model. This lexical/semantic/phonetic alignment approach yielded more than 160 million aligned entity pairs in 120 languages paired with English. Recognizing that each English is often aligned to mulitple entities in different target languages, we can join on English entities to obtain aligned entity pairs that directly pair two non-English entities (e.g., Arabic-French)