SiTNER: Improving Few-Shot Cross-lingual Nested Named Entity Recognition with high-quality pseudo-labels

Anonymous

Abstract:

Few-shot named entity recognition (NER) methods have shown preliminary effectiveness in flat tasks. However, existing methods still encounter difficulties when faced with cross-lingual and nested entity challenges due to the linguistic or nested structure gap. In this work, we propose a framework named SiTNER to deal with few-shot cross-lingual nested named entity recognition tasks. SiTNER mainly comprises two components: (1) contrastive span classification which could pull entities into corresponding prototype and generate high-quality pseudo-labels, and (2) masked pseudo data self-training which refine pseudo-labels and improves the span classification
via self-training strategy. We train SiTNER on the English dataset and evaluate it on the English, German, and Russian datasets, and experimental results show our method could get comparable results.

Paper Type: long

Research Area: Efficient/Low-Resource Methods for NLP

Contribution Types: Approaches to low-resource settings

Languages Studied: English, German, Russian

Consent To Share Submission Details: On behalf of all authors, we agree to the terms above to share our submission details.