Adversarial Transfer for Classical Chinese NER with Translation Word Segmentation

Abstract

AbstractClassical Chinese NER aims to automatically identify named entities in classical texts, which can effectively help people understand the content of classical texts. However, due to the difficulty of annotating classical Chinese texts, the scale of existing datasets seriously restricts the development of classical Chinese NER. To address this challenge, we propose an Adversarial Transfer for Classical Chinese NER (AT-CCNER) model, which transfers features learned from large-scale translation word segmentation to assist recognize classical Chinese named entities. In addition, to reduce the feature differences between modern and classical Chinese texts, AT-CCNER utilizes the adversarial method to better apply to classical Chinese texts. We experimentally demonstrate the effectiveness of our method on the open-source classical Chinese NER dataset C-CLUE. What’s more, we compare the effects of translation text of different scales on the experimental results. Our method improves Precision, Recall, and F1 by 3.61%, 3.45%, and 3.54%, respectively, compared to the BiLSTM-CRF model.KeywordsClassical Chinese NERChinese word segmentationAdversarial learningTransfer learning