Test Sets for Chinese Nonlocal Dependency Parsing

Test Sets for Chinese Nonlocal Dependency Parsing Manjuan Duan, William Schuler Department of Linguistics The Ohio State University 1712 Neil Ave. Columbus OH, 43210 {duan.59, schuler.77}@osu.edu Abstract Chinese is a language rich in nonlocal dependencies. Correctly resolving these dependencies is crucial in understanding the predicate- argument structure of a sentence. Making full use of the trace annotations in the Penn Chinese Treebank (Xue et al., 2005), this research contributes several test sets of Chinese nonlocal dependencies which occur in different grammatical constructions. These datasets can be used by an automatic dependency parser to evaluate its performance on nonlocal dependency resolution in various syntactic constructions in Chinese. Keywords: nonlocal dependency, parsing, Mandarin Chinese 1. Introduction Recovering unbounded dependencies is challenging in En- glish, as reported by Rimell et al. (2009). However, it serves as an important test of an automatic parser which cannot be easily accomplished by shallow language tech- niques. In spite of the low frequency of some nonlo- cal dependency constructions, correctly resolving these dependencies is crucial in understanding the underlying predicate-argument structure of a sentence. Although trace categories are annotated in the Penn Tree- bank (Xue et al., 2005), few state-of-the-art constituent parsers make use of these annotations to make predictions of the nonlocal dependencies. Categorial grammar annota- tions of the Treebank are advocated partly for their well- defined representations of filler-gap phenomena in natural languages. Parsers trained on categorial grammar annota- tions are found to produce superior performance in recover- ing nonlocal dependencies in English (Rimell et al., 2009; Nguyen et al., 2012). In this research we focus on various nonlocal dependency types in Mandarin Chinese, a language that makes heavy use of nonlocal dependencies (Kummerfeld et al., 2013). We make full use of the trace categories annotated in the Penn Chinese Treebank (Xue et al., 2005) to generate test sets for eight nonlocal dependency constructions. We evaluate the nonlocal dependency recovery performance of parsers trained on generalized categorial grammar annota- tions (Bach, 1981; Nguyen et al., 2012; Duan and Schuler, 2015) with the annotated test sets. We hope these test sets can make it easier to automatically evaluate nonlocal de- pendency recovery in Mandarin Chinese. 2. Nonlocal Dependency Data Sets 2.1. The Constructions In a nonlocal dependency, a constituent seems to be moved from its canonical position in the predicate-argument pat- tern, while it still needs to be interpreted in the position from which it is moved. For this research we examined the trace categories annotated in Penn Chinese Treebank and included those constructions where either both the trace and the filler were clearly annotated or they can reliably be re- covered from the syntactic trees. For example, the Chinese Penn Treebank has clear annotations about the locations of traces in relative clauses but the fillers are ‘WHNP’ empty categories. However, the head noun of a relative clause can always be reliably and accurately located given the tree structure. We therefore include relative clauses as a nonlo- cal construction in this study. The only exception is …

Content truncated to 500 words