L1-L2 Parallel Treebank of Learner Chinese: Overused and Underused Syntactic Structures

L1-L2 Parallel Treebank of Learner Chinese: Overused and Underused Syntactic Structures Keying Li, John Lee Department of Linguistics and Translation City University of Hong Kong keyingli3-c@my.cityu.edu.hk, jsylee@cityu.edu.hk Abstract We present a preliminary analysis on a corpus of texts written by learners of Chinese as a foreign language (CFL), annotated in the form of an L1-L2 parallel dependency treebank. The treebank consists of parse trees of sentences written by CFL learners (“L2 sentences”), parse trees of their target hypotheses (“L1 sentences”), and word alignment between the L1 sentences and L2 sentences. Currently, the treebank consists of 600 L2 sentences and 697 L1 sentences. We report the most overused and underused syntactic relations by the CFL learners, and discuss the underlying learner errors. Keywords: learner corpus, parallel treebank, Chinese as a foreign language 1. Introduction Learner corpora, which consist of texts written by non- native speakers, are increasingly used in quantitative studies in second language acquisition. Some of these corpora have been annotated to answer various research questions. To support analysis of grammatical mistakes made by learners, a number of them have been error-tagged (e.g., Yannakoudakis et al., 2011; Dahlmeier et al., 2013; Lee et al., 2016). To better characterize learner syntax, others have been part-of-speech (POS) tagged (e.g., Díaz- Negrillo et al., 2010; Reznicek et al., 2013), and syntactically analysed with constituent trees (e.g., Nagata and Sakaguchi, 2016) and dependency trees (e.g., Ragheb and Dickinson, 2014; Berzak et al., 2016). Building on learner treebanks, Lee et al. (2017b) proposed to use “L1-L2 parallel treebanks” — parse trees of non- native sentences (“L2 sentences”) aligned to their target hypotheses (“L1 sentences”) — to facilitate analyses of learner language. Figure 1 shows an example tree pair. It includes the parse tree of the learner sentence and of its target target hypothesis, both annotated in the Universal Dependencies (UD) scheme for Chinese (Leung et al., 2016; Lee et al., 2017), as well as word alignments between the two sentences. Such a treebank has the potential to enhance Contrastive Interlanguage Analysis (CIA) (Granger, 2015) and Error Analysis (EA) by supporting a greater range of automatic, quantitative studies. For CIA, they would enable comparisons between native and interlanguages not only on the lexical level but also on the syntactic level. For EA, parallel parse trees would give more fine-grained characterization of the syntactic environment in which learner errors occur. This paper reports on the construction of an L1-L2 parallel treebank for Chinese and presents a preliminary analysis. After summarizing previous work (Section 2), we give details on the texts in the treebank (Section 3) and on the linguistic annotations (Section 4). We then discuss the most overused and underused syntactic structures in the learner texts, as well as the underlying errors (Section 5). Finally, we conclude in Section 6. Figure 1: An example L1-L2 tree pair, including word alignments between the learner sentence (“L2”) and its target hypothesis (“L1”), and the parse trees of the two sentences, annotated in the Universal Dependencies scheme for Chinese (Leung …

Content truncated to 500 words