The contribution of error-tagged learner corpora to the assessment of language proficiency. Evidence from the International Corpus of Learner English

(eng) Learner corpora (LC) are electronic collections of spoken or written texts produced by foreign or second language learners. Despite its relative youth, this new resource displays considerable potential for interlanguage studies. It is, among others, currently being used to inform pedagogical tools (learner dictionaries, grammars, textbooks, CALL programs) and NLP tools such as spell- and grammar checkers. Another possible field which would most certainly gain from the use of LC is that of language testing, but as highlighted by Barker (2003), "they are not yet exploited fully by the language testing community". Our presentation will consider how a quantitative and qualitative analysis of the number and type of errors found in LC can contribute to the assessment of language proficiency levels. It focuses on learner texts extracted from the International Corpus of Learner English (Granger 2003) and which we error-tagged on the basis of the error tagging system designed at Louvain (Dagneaux et al 1998). Our aim was to establish clear links between certain error types and the different levels of proficiency found in the Common European Framework of Reference for Languages (CEF). As each of the error-tagged texts had already been assessed independently by a professional rater on the basis of the CEF guidelines, we were able to compare them with the level of proficiency each had been assigned independently. Our aim in doing so was to flesh out the descriptors found in the CEF which are very explicit on the learners' 'Can Dos' but only implicitly refer to their grammatical and lexical 'Can't Dos'.