Exploiting Lexical Dependencies from Large-Scale Data for Better Shift-Reduce Constituency Parsing

Proceedings of COLING 2012: Technical Papers, pages 3171–3186, COLING 2012, Mumbai, December 2012. Exploiting Lexical Dependencies from Large-Scale Data for Better Shift-Reduce Constituency Parsing Muhua ZHU Jingbo ZHU Huizhen WANG Natural Language Processing Laboratory Northeastern University, China zhumuhua@gmail.com, {zhujingbo,wanghuizhen}@mail.neu.edu.cn Abstract This paper proposes a method to improve shift-reduce constituency parsing by using lexical de- pendencies. The lexical dependency information is obtained from a large amount of auto-parsed data that is generated by a baseline shift-reduce parser on unlabeled data. We then incorporate a set of novel features defined on this information into the shift-reduce parsing model. The fea- tures can help to disambiguate action conflicts during decoding. Experimental results show that the new features achieve absolute improvements over a strong baseline by 0.9% and 1.1% on En- glish and Chinese respectively. Moreover, the improved parser outperforms all previously reported shift-reduce constituency parsers. Title and Abstract in Chinese 利 利 利用 用 用大 大 大规 规 规模 模 模数 数 数据 据 据词 词 词汇 汇 汇依 依 依存 存 存关 关 关系 系 系改 改 改进 进 进移 移 移进 进 进－ － －归 归 归约 约 约成 成 成分 分 分句 句 句法 法 法分 分 分析 析 析 本文提出了一种利用词汇依存关系改进移进－归约成分句法分析的方法。首先，我们利用 基准系统在大规模无标注数据上进行自动句法分析并从分析结果中抽取词汇依存关系。其 后，我们在词汇依存信息的基础上定义了一组新特征并将这些特征整合到移进－归约句法 分析模型 中。新特征用于帮助消除移进－归约过程中的动作歧义。实验结果表明，新特征 在英文和中文数据上分别取得了0.9% 和1.1%的性能改进。最终得到的句法分析器的性能 优于相关研究工作中所报告的移进－归约句法分析器的性能。 Keywords: Shift-reduce Constituency Parsing, Lexical Dependencies, Large-scale Data. Keywords in Chinese: 移进－归约成分句法分析，词汇依存，大规模数据. 3171 1 Introduction Due to the simplicity and running efficiency, shift-reduce parsing has been studied extensively for a variety of grammars, ranging from constituency parsing (Sagae and Lavie, 2005, 2006; Zhang and Clark, 2009) through dependency parsing (Nivre, 2004; Yamada and Matsumoto, 2003; Zhang and Clark, 2008) to CCG parsing (Zhang and Clark, 2011). In dependency and CCG parsing, shift-reduce parsing is among the best-performing algorithms (Huang and Sagae, 2010; Zhang and Clark, 2011). However, compared to commonly-used statistical parsers available on the web such as Charniak-Johnson (Charniak and Johnson, 2005) and Petrov-Klein (Petrov and Klein, 2007), shift-reduce constituency parsers still have room left for further improvements on parsing accuracy. There exist at least two major directions to advance shift-reduce constituency parsing. One direc- tion is to design better training and decoding algorithms. For example, in the respect of decod- ing, Sagae and Lavie (2006) proposed a best-first search strategy to expand the search space. In the respect of training, Zhang and Clark (2009) replaced local classifiers with a global learning algo- rithm. The other direction is to enrich feature representations for better shift-reduce constituency parsing, which will be the focus of this paper. In this direction, previous work has extensively studied a variety of features, all in the framework of supervised learning (Sagae and Lavie, 2005, 2006; Wang et al., 2006; Zhang and Clark, 2009). How to further enrich feature representations for better shift-reduce constituency parsing becomes a very challenging problem. In this paper, we solve this issue by using the information of lexi- cal head-modifier1 relations (a.k.a. lexical dependencies) (Collins, 1996). Previous work on other constituency parsers have shown the effectiveness of lexical dependency information on disam- biguating …

Content truncated to 500 words