Understanding Lexical Features for Chinese Essay Grading.

Essay grading is an important and difficult task in natural language processing. Most of the existing works focus on grading non-native English essays, such as essays in TOEFL. However, these works are not applicable for Chinese essays due to word segmentation and different syntax features. Considering lexical features are important for essay grading, in this paper, we study the expert evaluation standard and propose an interpretable lexical grading method for essays. We first study different levels of vocabulary provided by experts and introduce a quantitative evaluation framework on lexical features. Based on these standards, we quantify the Chinese essay dataset of 12 education grades in primary and middle schools and propose a set of interpretable features. Then a Bi-LSTM network model is proposed for semantically grading essay, which accepts a sequence of word vectors as input and integrates attention mechanism in terms of lexical richness. We evaluate our method on real datasets and the experimental results show that it outperforms other methods on the task of lexically Chinese essay grading. Besides, our method gives interpretable results, which are helpful for practical applications.