Evaluating performance of grammatical error detection to maximize learning effect
Ryo Nagata, Kazuhide Nakatani · International Conference on Computational Linguistics · 2010
This paper proposes a method for evaluating grammatical error detection methods to maximize the learning effect obtained by grammatical error detection. To achieve this, this paper sets out the following two hypotheses --- imperfect, rather than perfect, error detection maximizes learning effect; and precision-oriented error detection is better than a recall-oriented one in terms of learning effect. Experiments reveal that (i) precision-oriented error detection has a learning effect comparable to that of feedback by a human tutor, although the first hypothesis is not supported; (ii) precision-oriented error detection is better than recall-oriented in terms of learning effect; (iii) F-measure is not always the best way of evaluating error detection methods.