No abstract

In this paper we describe the CLEC corpus, an ongoing project set up at the University of Cádiz with the purpose of building up a large corpus of English as a 2L classified according to CEFR proficiency levels and formed to train statistical models for automatic proficiency assessment. The goal of this corpus is twofold: on the one hand it will be used as a data resource for the development of automatic text classification systems and, on the other, it has been used as a means of teaching innovation techniques.

The linguistic profiling of L2 learning texts can be taken as a model for automatic proficiency assessment of new texts. But proficiency levels are distinguished by many different linguistic features among which the use of cohesive devices can be a criterial element for level distinctions, either in the number of conjunctions used (quantitative) and/or in the type and variety of them (qualitative). We have carried such an analysis with a subgroup of the CLEC (CEFR-levelled English Corpus) using Coh-Metrix, a tool for computing computational cohesion and coherence metrics for written and spoken texts, but our results suggest that automatic proficiency level assessment needs a deeper examination of conjunctions that should rely on the analysis of conjunction-types use and conjunction varieties, with an analysis of lexical choice. A variable based on familiarity ranks could help to predict cohesive levels proficiencyoriented.