Exploring Adaptor Grammars for Native Language Identification

Proceedings of the 2012 Joint Conference on Empirical Methods in Natural Language Processing and Computational Natural Language Learning, pages 699–709, Jeju Island, Korea, 12–14 July 2012. c⃝2012 Association for Computational Linguistics Exploring Adaptor Grammars for Native Language Identification Sze-Meng Jojo Wong Mark Dras Centre for Language TechnologyMark Johnson Macquarie University Sydney, NSW, Australia {sze.wong,mark.dras,mark.johnson}@mq.edu.au Abstract The task of inferring the native language of an author based on texts written in a second language has generally been tackled as a clas- sification problem, typically using as features a mix of n-grams over characters and part of speech tags (for small and fixed n) and un- igram function words. To capture arbitrar- ily long n-grams that syntax-based approaches have suggested are useful, adaptor grammars have some promise. In this work we investi- gate their extension to identifying n-gram col- locations of arbitrary length over a mix of PoS tags and words, using both maxent and in- duced syntactic language model approaches to classification. After presenting a new, simple baseline, we show that learned collocations used as features in a maxent model perform better still, but that the story is more mixed for the syntactic language model. 1 Introduction The task of inferring the native language of an author based on texts written in a second language — na- tive language identification (NLI) — has, since the seminal work of Koppel et al. (2005), been primarily tackled as a text classification task using supervised machine learning techniques. Lexical features, such as function words, character n-grams, and part-of- speech (PoS) n-grams, have been proven to be use- ful in NLI (Koppel et al., 2005; Tsur and Rappoport, 2007; Estival et al., 2007). The recent work of Wong and Dras (2011), motivated by ideas from Second Language Acquisition (SLA), has shown that syn- tactic features — potentially capturing syntactic er- rors characteristic of a particular native language — improve performance over purely lexical ones. PoS n-grams can be leveraged to characterise sur- face syntactic structures: in Koppel et al. (2005), for example, ungrammatical structures were approx- imated by rare PoS bigrams. For the purpose of NLI, small n-gram sizes like bigram or trigram might not suffice to capture sequences that are characteristic of a particular native language. On the other hand, an attempt to represent these with larger n-grams would not just lead to feature sparsity problems, but also computational efficiency issues. Some form of fea- ture selection should then come into play. Adaptor grammars (Johnson, 2010), a hierarchi- cal non-parametric extension of PCFGs (and also in- terpretable as an extension of LDA-based topic mod- els), hold out some promise here. In that initial work, Johnson’s model learnt collocations of arbi- trary length such as gradient descent and cost func- tion, under a topic associated with machine learning. Hardisty et al. (2010) applied this idea to perspective classification, learning collocations such as pales- tinian violence and palestinian freedom, the use of which as features was demonstrated to help the clas- sification of texts from the …

Content truncated to 500 words