Joint POS Tagging and Transition-based Constituent Parsing in Chinese with Non-local Features

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 733–742, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Joint POS Tagging and Transition-based Constituent Parsing in Chinese with Non-local Features Zhiguo Wang Brandeis University Waltham, MA, USA zgwang@brandeis.edu Nianwen Xue Brandeis University Waltham, MA, USA xuen@brandeis.edu Abstract We propose three improvements to ad- dress the drawbacks of state-of-the-art transition-based constituent parsers. First, to resolve the error propagation problem of the traditional pipeline approach, we incorporate POS tagging into the syntac- tic parsing process. Second, to allevi- ate the negative influence of size differ- ences among competing action sequences, we align parser states during beam-search decoding. Third, to enhance the pow- er of parsing models, we enlarge the fea- ture set with non-local features and semi- supervised word cluster features. Exper- imental results show that these modifica- tions improve parsing performance signif- icantly. Evaluated on the Chinese Tree- Bank (CTB), our final performance reach- es 86.3% (F1) when trained on CTB 5.1, and 87.1% when trained on CTB 6.0, and these results outperform all state-of-the-art parsers. 1 Introduction Constituent parsing is one of the most fundamen- tal tasks in Natural Language Processing (NLP). It seeks to uncover the underlying recursive phrase structure of sentences. Most of the state-of-the- art parsers are based on the PCFG paradigm and chart-based decoding algorithms (Collins, 1999; Charniak, 2000; Petrov et al., 2006). Chart-based parsers perform exhaustive search with dynam- ic programming, which contributes to their high accuracy, but they also suffer from higher run- time complexity and can only exploit simple local structural information. Transition-based constituent parsing (Sagae and Lavie, 2005; Wang et al., 2006; Zhang and Clark, 2009) is an attractive alternative. It utilizes a se- ries of deterministic shift-reduce decisions to con- struct syntactic trees. Therefore, it runs in linear time and can take advantage of arbitrarily complex structural features from already constructed sub- trees. The downside is that they only search a tiny fraction of the whole space and are therefore com- monly considered to be less accurate than chart- based parsers. Recent studies (Zhu et al., 2013; Zhang et al., 2013) show, however, that this ap- proach can also achieve the state-of-the-art perfor- mance with improved training procedures and the use of additional source of information as features. However, there is still room for improvemen- t for these state-of-the-art transition-based con- stituent parsers. First, POS tagging is typically performed separately as a preliminary step, and POS tagging errors will propagate to the parsing process. This problem is especially severe for lan- guages where the POS tagging accuracy is rela- tively low, and this is the case for Chinese where there are fewer contextual clues that can be used to inform the tagging process and some of the tagging decisions are actually influenced by the syntactic structure of the sentence. This creates a chicken and egg problem that needs to be ad- dressed when designing a parsing model. Second, due to the existence of unary …

Content truncated to 500 words