ZEN: Pre-training Chinese Text Encoder Enhanced by N-gram Representations

The pre-training of text encoders normally processes text as a sequence of tokens corresponding to small text units, such as word pieces in English and characters in Chinese. It omits information carried by larger text granularity, and thus the encoders cannot easily adapt to certain combinations of characters. This leads to a loss of important semantic information, which is especially problematic for Chinese because the language does not have explicit word boundaries. In this paper, we propose ZEN, a BERT-based Chinese (Z) text encoder Enhanced by N-gram representations, where different combinations of characters are considered during training. As a result, potential word or phase boundaries are explicitly pre-trained and fine-tuned with the character encoder (BERT). Therefore ZEN incorporates the comprehensive information of both the character sequence and words or phrases it contains. Experimental results illustrated the effectiveness of ZEN on a series of Chinese NLP tasks. We show that ZEN, using less resource than other published encoders, can achieve state-of-the-art performance on most tasks. Moreover, it is shown that reasonable performance can be obtained when ZEN is trained on a small corpus, which is important for applying pre-training techniques to scenarios with limited data. The code and pre-trained models of ZEN are available at https://github.com/sinovation/zen.

PDF Abstract Findings of 2020 PDF Findings of 2020 Abstract
Task Dataset Model Metric Name Metric Value Global Rank Result Benchmark
Chinese Sentiment Analysis ChnSentiCorp ZEN (Init with Chinese BERT) F1 96.08 # 1
Chinese Sentiment Analysis ChnSentiCorp ZEN (Random Init) F1 94.42 # 2
Chinese Sentiment Analysis ChnSentiCorp Dev ZEN (Random Init) F1 94.87 # 2
Chinese Sentiment Analysis ChnSentiCorp Dev ZEN (Init with Chinese BERT) F1 95.66 # 1
Chinese Part-of-Speech Tagging CTB5 ZEN (Init with Chinese BERT) F1 96.64 # 1
Chinese Part-of-Speech Tagging CTB5 ZEN (Random Init) F1 95.82 # 3
Chinese Part-of-Speech Tagging CTB5 Dev ZEN (Init with Chinese BERT) F1 97.43 # 1
Chinese Part-of-Speech Tagging CTB5 Dev ZEN (Random Init) F1 96.12 # 2
Chinese Sentence Pair Classification LCQMC ZEN (Random Init) F1 85.27 # 4
Chinese Sentence Pair Classification LCQMC ZEN (Init with Chinese BERT) F1 87.95 # 2
Chinese Sentence Pair Classification LCQMC Dev ZEN (Random Init) F1 88.1 # 3
Chinese Sentence Pair Classification LCQMC Dev ZEN (Init with Chinese BERT) F1 90.2 # 2
Chinese Word Segmentation MSR ZEN (Init with Chinese BERT) F1 98.35 # 4
Chinese Word Segmentation MSR ZEN (Random Init) F1 97.89 # 6
Chinese Named Entity Recognition MSRA ZEN (Init with Chinese BERT) F1 95.25 # 9
Chinese Named Entity Recognition MSRA ZEN (Random Init) F1 93.24 # 18
Chinese Document Classification THUCNews ZEN (Init with Chinese BERT) F1 97.64 # 2
Chinese Document Classification THUCNews ZEN (Random Init) F1 96.87 # 3
Chinese Document Classification THUCNews Dev ZEN (Random Init) F1 97.2 # 3
Chinese Document Classification THUCNews Dev ZEN (Init with Chinese BERT) F1 97.66 # 2
Chinese Sentence Pair Classification XNLI ZEN (Init with Chinese BERT) F1 79.2 # 2
Chinese Sentence Pair Classification XNLI ZEN (Random Init) F1 77.03 # 3
Chinese Sentence Pair Classification XNLI Dev ZEN (Init with Chinese BERT) F1 80.48 # 2
Chinese Sentence Pair Classification XNLI Dev ZEN (Random Init) F1 77.11 # 3

Methods


No methods listed for this paper. Add relevant methods here