A Comparative Study of Feature Representations for Spanish L2 Text Complexity Across Pedagogical and Authentic Corpora

Resumen

This study investigates the effectiveness of three feature representations for Spanish L2 CEFR level prediction: traditional readability metrics, CEFR-based lexical overlap features, and unigram TF-IDF. We evaluate these features individually and in combination under varying data conditions across two types of datasets: an authentic learner corpus (CAES) and two pedagogical datasets (HablaCultura and Kwiziq). Results show that distributional features achieve the highest performance when sufficient training data is available. In contrast, in smaller or cross-dataset settings, traditional readability metrics combined with CEFR-based lexical features provide more stable performance. We also observe limited generalization across dataset types, particularly when transferring from learner-produced to pedagogical texts. Overall, the findings highlight the need to adapt feature representations to data conditions and domain characteristics for Spanish L2 CEFR prediction.