On the Automatic Analysis of Learner Language. Introduction to the Special Issue

Natural language processing (NLP) has long been used to automatically analyze language produced by language learners, typically aimed at providing individualized feedback and learner modeling in Intelligent Computer-Assisted Language Learning systems (cf. Heift & Schulze 2007). While much interesting research has been reported, it is difficult to determine the state of the art for the automatic analysis of learner language. Which error types and other learner language properties can be detected and diagnosed automatically? How reliably can this be done, for which kind of learner language, resulting from which types of tasks? For sustainable progress on the automatic analysis of learner language it arguably is crucial to answer these questions, to discuss and compare the performance of different analysis methods on real-life learner data sets. As an essential prerequisite for addressing these issues, it is necessary to determine which learner language properties are useful or important to analyze in order to provide feedback and model language acquisition – a question which highlights the need for an intensive interdisciplinary dialogue between the fields of Intelligent Computer-Assisted Language Learning (ICALL), Second Language Acquisition (SLA), and Foreign Language Teaching (FLT). Relatedly, the questions arising for the automatic analysis of learner language in ICALL intersect in important ways with research on learner corpora (cf. Granger 1998). Learner corpora in principle can help validate generalizations about language acquisition and provide a broad empirical basis for the development of new hypotheses and theories in SLA. However, to find the relevant classes of examples, the linguistic terminology used to single out the learner language aspects of interest needs to be mapped to instances in the corpus. Effective querying of corpora thus often requires reference to annotated linguistic abstractions instead of extensionally characterizing strings (cf. Meurers & Müller 2007). Most of the work on annotating learner corpora has focused on the annotation of learner errors, for which a number of annotation schemes have been developed (Dı́az-Negrillo & Fernández-Domı́nguez 2006). Yet, parallel to the ICALL situation mentioned above, there seems to be no agreement as to which distinctions are