Revisiting Pre-Trained Models for Chinese Natural Language Processing

By Yiming Cui, Wanxiang Che, Ting-Wei Liu, Bing Qin, Shijin Wang, Guoping Hu. 2020

878 citations recorded by the source corpus.

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. https://github.com/ymcui/MacBERT

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Large Language Models, Natural Language Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

1. Ting-Wei Liu
2. Wanxiang Che
3. Yiming Cui
4. Hua Wu
5. Shijin Wang
6. Guoping Hu
7. Haifeng Wang
8. Ziqing Yang

In Computer Vision and Pattern Recognition

1. Ting-Wei Liu
2. Yiming Cui
3. Shijin Wang
4. Wanxiang Che
5. Guoping Hu
6. Haifeng Wang
7. Hua Wu
8. Jingren Zhou

In Large Language Models

1. Ziqing Yang
2. Haifeng Wang
3. Yiming Cui
4. Hua Wu
5. An Yang
6. Ming Ding
7. Jingren Zhou
8. Yukun Li

In Natural Language Processing

1. Ting-Wei Liu
2. Hua Wu
3. Yiming Cui
4. Guoping Hu
5. Shijin Wang
6. Wanxiang Che
7. Haifeng Wang
8. Zhipeng Chen

Citations in the corpus

• M6: A Chinese Multimodal Pretrainer (2021, background): “[5] employs unreleased extended data that are 10 times larger than the CN-Wikipedia to pretrain their Chinese language model.”
• ERNIE 3.0 Titan: Exploring Larger-scale Knowledge Enhanced Pre-training for Language Understanding and Generation (2021, neutral): “0 [52], SKEP [93], RoBERTa-wwm-ext-large [94] (marked as RoBERTa*), ALBERT [95], MacBERT [96], Zen 2.”
• SPARTA: Efficient Open-Domain Question Answering via Sparse Transformer Matching Retrieval (2020, methodology): “For Chinese tasks, we fine-tune on chinese-bert-wwm (Cui et al., 2020).”
• FewCLUE: A Chinese Few-shot Learning Evaluation Benchmark (2021, methodology): “Our default baseline model, RoBERTa, is implemented using RoBERTa-wwm-ext (Cui et al., 2020) 2 .”
• Towards artificial general intelligence via a multimodal foundation model (2021, methodology): “, RoBERTa [42]) is then used to map text tokens to a sequence of feature vectors (each feature vector corresponds to a word).”
• A Sentence Cloze Dataset for Chinese Machine Reading Comprehension (2020, methodology): “2 Experimental Setups We adopt Chinese BERT-base, BERT-multilingual, Chinese BERT-wwmand RoBERTawith whole word masking (Cui et al., 2019a; Cui et al., 2020) as backbones.”
• ERNIE-Gram: Pre-Training with Explicitly N-Gram Masked Language Modeling for Natural Language Understanding (2020, methodology): “To inject n-gram information into pre-training, many works (Sun et al., 2019b; Cui et al., 2019; Joshi et al., 2020) extend BERT’s masked language modeling (MLM) from masking individual tokens to contiguous sequences of n tokens.”