Exploiting Web-Derived Selectional Preference to Improve Statistical Dependency Parsing

Proceedings of the 49th Annual Meeting of the Association for Computational Linguistics, pages 1556–1565, Portland, Oregon, June 19-24, 2011. c⃝2011 Association for Computational Linguistics Exploiting Web-Derived Selectional Preference to Improve Statistical Dependency Parsing Guangyou Zhou, Jun Zhao∗, Kang Liu, and Li Cai National Laboratory of Pattern Recognition Institute of Automation, Chinese Academy of Sciences 95 Zhongguancun East Road, Beijing 100190, China {gyzhou,jzhao,kliu,lcai}@nlpr.ia.ac.cn Abstract In this paper, we present a novel approach which incorporates the web-derived selec- tional preferences to improve statistical de- pendency parsing. Conventional selectional preference learning methods have usually fo- cused on word-to-class relations, e.g., a verb selects as its subject a given nominal class. This paper extends previous work to word- to-word selectional preferences by using web- scale data. Experiments show that web-scale data improves statistical dependency pars- ing, particularly for long dependency relation- ships. There is no data like more data, perfor- mance improves log-linearly with the number of parameters (unique N-grams). More impor- tantly, when operating on new domains, we show that using web-derived selectional pref- erences is essential for achieving robust per- formance. 1 Introduction Dependency parsing is the task of building depen- dency links between words in a sentence, which has recently gained a wide interest in the natural lan- guage processing community. With the availabil- ity of large-scale annotated corpora such as Penn Treebank (Marcus et al., 1993), it is easy to train a high-performance dependency parser using super- vised learning methods. However, current state-of-the-art statistical de- pendency parsers (McDonald et al., 2005; McDon- ald and Pereira, 2006; Hall et al., 2006) tend to have ∗Correspondence author: jzhao@nlpr.ia.ac.cn lower accuracies for longer dependencies (McDon- ald and Nivre, 2007). The length of a dependency from word wi to word wj is simply equal to |i − j|. Longer dependencies typically represent the mod- ifier of the root or the main verb, internal depen- dencies of longer NPs or PP-attachment in a sen- tence. Figure 1 shows the F1 score1 relative to the dependency length on the development set by using the graph-based dependency parsers (McDonald et al., 2005; McDonald and Pereira, 2006). We note that the parsers provide very good results for adja- cent dependencies (96.89% for dependency length =1), while the dependency length increases, the ac- curacies degrade sharply. These longer dependen- cies are therefore a major opportunity to improve the overall performance of dependency parsing. Usu- ally, these longer dependencies can be parsed de- pendent on the specific words involved due to the limited range of features (e.g., a verb and its mod- ifiers). Lexical statistics are therefore needed for resolving ambiguous relationships, yet the lexical- ized statistics are sparse and difficult to estimate di- rectly. To solve this problem, some information with different granularity has been investigated. Koo et al. (2008) proposed a semi-supervised dependency parsing by introducing lexical intermediaries at a coarser level than words themselves via a cluster method. This approach, however, ignores the se- lectional preference for word-to-word interactions, such as head-modifier relationship. Extra resources 1Precision represents …

Content truncated to 500 words