2019 Volume 22 Pages 23-43
Over the last four decades, the constructs of complexity, accuracy, and fluency have been in focus in the analysis of language learners’ performance. However, due to the polysemous nature of complexity, more and more sub-constructs have been assumed, making holistic measurement difficult. This study aims to construct a more appropriate measurement model of L2 complexity by implementing finer-grained and relatively novel linguistic indices for capturing subordinate constructs that could not be measured by conventional indices. By utilizing five natural language processing tools, conventional and fine-grained indices of complexity were computed from 503 argumentative essays written by Japanese English learners. First, exploratory factor analysis was performed on linguistic index values and the extracted factor structures behind them. Second, confirmatory factor analysis was conducted to confirm whether the structure fits the data. Finally, a structural equation model of complexity constructs to predict essay scores was tested to evaluate its applicability to writing evaluation. The result of a series of factor analyses showed that the extracted factor structures reasonably fitted to the data for syntactic complexity (CFI = .901 and RMSEA = .071) and for lexical complexity (CFI = .978 and RMSEA = .051). Furthermore, the result of Structural Equation Modeling (SEM) analysis, which was proposed as a predictive model, accounted for 32.3 % of the variance of essay scores (CFI = .916 and RMSEA = .077). Overall, the findings showed the effectiveness of the proposed approach, which combined conventional linguistic features with fine-grained and relatively novel indices.