Joint Chinese Word Segmentation, POS Tagging and Parsing

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 501–511, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Joint Chinese Word Segmentation, POS Tagging and Parsing Xian Qian Yang Liu Computer Science Department The University of Texas at Dallas qx,yangl@hlt.utdallas.edu Abstract In this paper, we propose a novel decoding al- gorithm for discriminative joint Chinese word segmentation, part-of-speech (POS) tagging, and parsing. Previous work often used a pipeline method – Chinese word segmentation followed by POS tagging and parsing, which suffers from error propagation and is unable to leverage information in later modules for earlier components. In our approach, we train the three individual models separately during training, and incorporate them together in a u- nified framework during decoding. We extend the CYK parsing algorithm so that it can deal with word segmentation and POS tagging fea- tures. As far as we know, this is the first work on joint Chinese word segmentation, POS tag- ging and parsing. Our experimental result- s on Chinese Tree Bank 5 corpus show that our approach outperforms the state-of-the-art pipeline system. 1 Introduction For Asian languages such as Japanese and Chi- nese that do not contain explicitly marked word boundaries, word segmentation is an important first step for many subsequent language processing tasks, such as POS tagging, parsing, semantic role label- ing, and various applications. Previous studies for POS tagging and syntax parsing on these languages sometimes assume that gold standard word segmen- tation information is provided, which is not the re- al scenario. In a fully automatic system, a pipeline approach is often adopted, where raw sentences are first segmented into word sequences, then POS tag- ging and parsing are performed. This kind of ap- proach suffers from error propagation. For exam- ple, word segmentation errors will result in tagging and parsing errors. Additionally, early modules can- not use information from subsequent modules. In- tuitively a joint model that performs the three tasks together should help the system make the best deci- sions. In this paper, we propose a unified model for joint Chinese word segmentation, POS tagging, and pars- ing. Three sub-models are independently trained using the state-of-the-art methods. We do not use the joint inference algorithm for training because of the high complexity caused by the large amount of parameters. We use linear chain Conditional Ran- dom Fields (CRFs) (Lafferty et al., 2001) to train the word segmentation model and POS tagging model, and averaged perceptron (Collins, 2002) to learn the parsing model. During decoding, parameters of each sub-model are scaled to represent its importance in the joint model. Our decoding algorithm is an exten- sion of CYK parsing. Initially, weights of all possi- ble words together with their POS tags are calcu- lated. When searching the parse tree, the word and POS tagging features are dynamically generated and the transition information of POS tagging is consid- ered in the span merge operation. Experiments are conducted on Chinese …

Content truncated to 500 words