Mcl-Ner: Cross-Lingual Named Entity Recognition Via Multi-view Contrastive Learning

Cross-lingual named entity recognition (CrossNER) faces challenges stemmingfrom uneven performance due to the scarcity of multilingual corpora, especiallyfor non-English data. While prior efforts mainly focus on data-driven transfermethods, a significant aspect that has not been fully explored is aligning bothsemantic and token-level representations across diverse languages. In thispaper, we propose Multi-view Contrastive Learning for Cross-lingual NamedEntity Recognition (mCL-NER). Specifically, we reframe the CrossNER task into aproblem of recognizing relationships between pairs of tokens. This approachtaps into the inherent contextual nuances of token-to-token connections withinentities, allowing us to align representations across different languages. Amulti-view contrastive learning framework is introduced to encompass semanticcontrasts between source, codeswitched, and target sentences, as well ascontrasts among token-to-token relations. By enforcing agreement within bothsemantic and relational spaces, we minimize the gap between source sentencesand their counterparts of both codeswitched and target sentences. Thisalignment extends to the relationships between diverse tokens, enhancing theprojection of entities across languages. We further augment CrossNER bycombining self-training with labeled source data and unlabeled target data. Ourexperiments on the XTREME benchmark, spanning 40 languages, demonstrate thesuperiority of mCL-NER over prior data-driven and model-based approaches. Itachieves a substantial increase of nearly +2.0 F_1 scores across a broadspectrum and establishes itself as the new state-of-the-art performer.