Revisiting Pre-Trained Models for Chinese Natural Language Processing
Yiming Cui, Wanxiang Che, Ting Liu, Bing Qin, Shijin Wang, Guoping Hu
TL;DR
The paper revisits Chinese pre-trained language models, adapting English-centric techniques to Chinese and introducing MacBERT, which reframes MLM as a correction task using similar-word replacements and enhanced masking strategies. By leveraging N-gram masking, WWM, and SOP, MacBERT achieves state-of-the-art results on several Chinese NLP benchmarks and demonstrates the importance of MLM design over NSP-type objectives in Chinese. Extensive experiments across eight Chinese datasets show robust gains on machine reading comprehension and competitive performance on classification tasks. The authors also release a Chinese pre-trained language model series to accelerate research and provide ablations to dissect the contributions of each component.
Abstract
Bidirectional Encoder Representations from Transformers (BERT) has shown marvelous improvements across various NLP tasks, and consecutive variants have been proposed to further improve the performance of the pre-trained language models. In this paper, we target on revisiting Chinese pre-trained language models to examine their effectiveness in a non-English language and release the Chinese pre-trained language model series to the community. We also propose a simple but effective model called MacBERT, which improves upon RoBERTa in several ways, especially the masking strategy that adopts MLM as correction (Mac). We carried out extensive experiments on eight Chinese NLP tasks to revisit the existing pre-trained language models as well as the proposed MacBERT. Experimental results show that MacBERT could achieve state-of-the-art performances on many NLP tasks, and we also ablate details with several findings that may help future research. Resources available: https://github.com/ymcui/MacBERT