Zero-Shot Entity Recognition via Multi-Source Projection and Unlabeled Data

Abstract

With the widespread of transfer learning, great success has been achieved in named entity recognition on languages without labeled data. Prior works can be mainly classified into two categories: direct model transfer and annotation projection. However, these works only utilize part of available data, including monolingual word embeddings, labeled data in multiple source languages, and unlabeled data in the target language. To make full use of these data and extract more information, we propose to train multiple teacher models on multiple source languages via the annotation projection method. Then a student model is trained on the unlabeled data in the target language with supervision from all labels obtained from multiple teacher models. Extensive experiments are conducted on three benchmark datasets, and the experimental results prove the effectiveness of the proposed method.