Adversarial Training Lattice LSTM for Chinese Named Entity Recognition of Social Media Texts
Runmei Zhang, Hai Huang, Lei Yin, Weiwei Cao · 2023
Chinese named entity recognition in social media is a challenging task. Existing methods based on word-level information or external knowledge are affected by Chinese word segmentation errors, insufficient exploration of potential words, and a lack of training data. Although most of the pre-training takes into account the semantic information of the word context to make it characterize multiple meanings of a word; however, excessive focus on the contextual information can result in inconsistent labeling of the same entity as the contribution of different words to entity recognition is not necessarily positive. To address the above problems, a model combining an improved generative adversarial network with Lattice LSTM-Attention that uses global feature vectors is proposed. The dual-layer BiLSTM serves as the generator model in the GAN, integrating positive samples consistent with expert-annotated data from human-annotated datasets, thus resolving the issue of limited training data. Additionally, a character glyph embedding layer is added to the pre-trained RoBERTa to discover more potential words. A Lattice LSTM-Attention sub-model with global feature representation is introduced, which achieves good results in Chinese word segmentation and solves the problem of inconsistent labeling of the same entity. Experimental results on the manually labeled Weibo2022 hot search dataset show that the model achieves more desirable results compared to other modeling approaches of its kind.