NEB-Filter: A Simple but Effective Filter based on Named Entity Boundaries for Low-Resource Cross-Lingual NER

Named Entity Recognition (NER) on low-resource languages remains a challenging task due to the scarcity of labeled data. With the advent of large-scale neural machine translation systems and multilingual pre-training models, it has become possible to transfer knowledge from the resource-rich source language side (e.g., English) to the resource-poor target language side. In this paper, we focus on scenarios where the source language side labeled data is also insufficient. We found that transfer entity boundary information alone cross the languages is much easier than transfer the whole labeled information, which includes not only the entity boundary but also the entity category. Therefore, we propose a simple but effective cross-lingual NER method called NEB-Filter for the first time. NEB-Filter trains a filter based on the boundaries of named entities, which can greatly improve the precision of the cross-lingual NER model, but only brings a small reduction in recall. Moreover, our method has been shown to benefit from knowledge distillation (KD), leading to greater improvements on F1 score. A series of experiments have shown encouraging results for our approach in low-resource cross-lingual NER.