[Paper Review] PKUSEG: A Toolkit for Multi-Domain Chinese Word Segmentation

Ruixuan Luo, Jingjing Xu+4 more|arXiv (Cornell University)|Jun 27, 2019

Natural Language Processing Techniques10 references108 citations

TL;DR

PKUSEG provides a multi-domain Chinese word segmentation toolkit with domain-specific models, unsupervised domain adaptation, and POS tagging, aiming for high out-of-the-box performance and easy transfer learning.

ABSTRACT

Chinese word segmentation (CWS) is a fundamental step of Chinese natural language processing. In this paper, we build a new toolkit, named PKUSEG, for multi-domain word segmentation. Unlike existing single-model toolkits, PKUSEG targets multi-domain word segmentation and provides separate models for different domains, such as web, medicine, and tourism. Besides, due to the lack of labeled data in many domains, we propose a domain adaptation paradigm to introduce cross-domain semantic knowledge via a translation system. Through this method, we generate synthetic data using a large amount of unlabeled data in the target domain and then obtain a word segmentation model for the target domain. We also further refine the performance of the default model with the help of synthetic data. Experiments show that PKUSEG achieves high performance on multiple domains. The new toolkit also supports POS tagging and model training to adapt to various application scenarios. The toolkit is now freely and publicly available for the usage of research and industry.

Motivation & Objective

• Motivate robust Chinese word segmentation across diverse domains (news, web, medicine, tourism).
• Provide domain-specific pre-trained models and a coarse-to-fine fine-tuning strategy.
• Introduce unsupervised domain adaptation to leverage unlabeled data from target domains.
• Enable POS tagging and user-trained models to support varied downstream tasks.

Proposed method

• Use a fast, high-precision CRF model with Viterbi inference for segmentation.
• Apply adaptive online gradient descent (ADF) to train high-dimensional CRF features efficiently.
• Pre-train a coarse-grained model on mixed-domain data (news and web, with CTB) and fine-tune on domain data.
• Introduce masked Attention Augmentation (mA^2) to derive synthetic data via translation and a Transformer, enabling unsupervised domain adaptation.
• Construct a large-scale domain vocabulary (~850K words) from multiple sources to improve coverage.
• Provide tooling for domain-specific and coarse-grained segmentation, as well as user-defined dictionaries and training of new models.

Experimental results

Research questions

• RQ1 Can PKUSEG deliver high segmentation accuracy across multiple domains with domain-specific models?
• RQ2 Does pre-training on mixed-domain data improve performance on domain-specific targets?
• RQ3 Can unsupervised domain adaptation (mA^2) generate useful target-domain data from unlabeled resources?
• RQ4 How does the refined default model generalize to out-of-domain data compared to domain-specific models?

Key findings

• Domain-specific models outperform the default model in all tested domains.
• Pre-training plus fine-tuning yields higher average F1 across medicines, web, news, and tourism.
• Unsupervised domain adaptation (mA^2) improves domain performance over the default when labeled data are scarce (average gains observed in Science, Art, Entertainment).
• Refining the default model with synthetic data improves out-of-domain performance (+0.67 F1 on out-of-domain datasets) without hurting in-domain performance.
• A large-scale domain vocabulary substantially boosts coverage for domain-specific terms (total ~850K words).
• POS tagging is supported alongside segmentation to broaden applicability.