Adversarial Multi-Criteria Learning for Chinese Word Segmentation

Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1193–1203 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1110 Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1193–1203 Vancouver, Canada, July 30 - August 4, 2017. c⃝2017 Association for Computational Linguistics https://doi.org/10.18653/v1/P17-1110 Adversarial Multi-Criteria Learning for Chinese Word Segmentation Xinchi Chen, Zhan Shi, Xipeng Qiu∗, Xuanjing Huang Shanghai Key Laboratory of Intelligent Information Processing, Fudan University School of Computer Science, Fudan University 825 Zhangheng Road, Shanghai, China {xinchichen13,zshi16,xpqiu,xjhuang}@fudan.edu.cn Abstract Different linguistic perspectives causes many diverse segmentation criteria for Chinese word segmentation (CWS). Most existing methods focus on improve the per- formance for each single criterion. Howe- ver, it is interesting to exploit these dif- ferent criteria and mining their common underlying knowledge. In this paper, we propose adversarial multi-criteria learning for CWS by integrating shared knowledge from multiple heterogeneous segmentation criteria. Experiments on eight corpora with heterogeneous segmentation criteria show that the performance of each corpus obtains a significant improvement, compa- red to single-criterion learning. Source co- des of this paper are available on Github1. 1 Introduction Chinese word segmentation (CWS) is a prelimi- nary and important task for Chinese natural lan- guage processing (NLP). Currently, the state-of- the-art methods are based on statistical supervi- sed learning algorithms, and rely on a large-scale annotated corpus whose cost is extremely expen- sive. Although there have been great achievements in building CWS corpora, they are somewhat in- compatible due to different segmentation criteria. As shown in Table 1, given a sentence “姚明进 入总决赛 (YaoMing reaches the final)”, the two commonly-used corpora, PKU’s People’s Daily (PKU) (Yu et al., 2001) and Penn Chinese Tree- bank (CTB) (Fei, 2000), use different segmenta- tion criteria. In a sense, it is a waste of resources if we fail to fully exploit these corpora. ∗Corresponding author. 1https://github.com/FudanNLP Corpora Yao Ming reaches the final CTB 姚明 进入 总决赛 PKU 姚 明 进入 总 决赛 Table 1: Illustration of the different segmentation criteria. Recently, some efforts have been made to ex- ploit heterogeneous annotation data for Chinese word segmentation or part-of-speech tagging (Ji- ang et al., 2009; Sun and Wan, 2012; Qiu et al., 2013; Li et al., 2015, 2016). These methods adop- ted stacking or multi-task architectures and sho- wed that heterogeneous corpora can help each ot- her. However, most of these model adopt the shal- low linear classifier with discrete features, which makes it difficult to design the shared feature spa- ces, usually resulting in a complex model. Fortu- nately, recent deep neural models provide a con- venient way to share information among multiple tasks (Collobert and Weston, 2008; Luong et al., 2015; Chen et al., 2016). In this paper, we propose an adversarial multi- criteria learning for CWS by integrating shared knowledge from multiple segmentation criteria. Specifically, we regard each segmentation cri- terion as a single task and propose three diffe- rent shared-private models under the framework of multi-task learning (Caruana, 1997; …

Content truncated to 500 words