Parse Imputation for Dependency Annotations

Proceedings of the 53rd Annual Meeting of the Association for Computational Linguistics and the 7th International Joint Conference on Natural Language Processing, pages 1385–1394, Beijing, China, July 26-31, 2015. c⃝2015 Association for Computational Linguistics Parse Imputation for Dependency Annotations Jason Mielens1 Liang Sun2 1Department of Linguistics 2Department of Mechanical Engineering The University of Texas at Austin The University of Texas at Austin {jmielens,jbaldrid}@utexas.edu sally722@utexas.edu Jason Baldridge1 Abstract Syntactic annotation is a hard task, but it can be made easier by allowing annotators flexibility to leave aspects of a sentence underspecified. Unfortunately, partial an- notations are not typically directly usable for training parsers. We describe a method for imputing missing dependencies from sentences that have been partially anno- tated using the Graph Fragment Language, such that a standard dependency parser can then be trained on all annotations. We show that this strategy improves perfor- mance over not using partial annotations for English, Chinese, Portuguese and Kin- yarwanda, and that performance competi- tive with state-of-the-art unsupervised and weakly-supervised parsers can be reached with just a few hours of annotation. 1 Introduction Linguistically annotated data is produced for many purposes in many contexts. It typically requires considerable effort, particularly for lan- guage documentation efforts in which tooling, data, and expertise in the language are scarce. The challenge presented by this scarcity is com- pounded when doing deeper analysis, such as syn- tactic structure, which typically requires greater expertise and existing tooling. In such scenar- ios, unsupervised approaches are a tempting strat- egy. While the performance of unsupervised dependency parsing has improved greatly since Klein and Manning’s (2004) Dependency Model with Valence (DMV), state-of-the-art unsuper- vised parsers still perform well below supervised approaches (Martins et al., 2010; Spitkovsky et al., 2012; Blunsom and Cohn, 2010). Additionally, they typically require large amounts of raw data. While this is not a problem for some languages, many of the world’s languages do not have a clean, digitized corpus available.1 For instance, the ap- proach of Naseem et al. (2010) is unsupervised in the sense that it requires no dependency annota- tions, but it still makes use of the raw version of the full Penn Treebank. The approach of Mare- cek et al. (2013) requires extra unlabeled texts to estimate parameters. Another strategy is to exploit small amounts of supervision or knowledge. Naseem et al. (2010) use a set of universal dependency rules and obtain substantial gains over unsupervised methods in many languages. Spitkovsky et al. (2010b; 2011) use web mark-up and punctuation as additional an- notations. Alternatively, one could try to obtain actual dependency annotations cheaply. We use the Graph Fragment Language (GFL), which was created with the goal of making annotations eas- ier for experts and possible for novices (Schneider et al., 2013; Mordowanec et al., 2014). GFL sup- ports partial annotations, so annotators can omit obvious dependencies or skip difficult construc- tions. The ability to focus on portions of a sen- tence frees the annotator to target constituents and dependencies that maximize information that will …

Content truncated to 500 words