自動建構中文作文評分系統：產生、篩選與評估

Abstract

在本論文中，我們探討概念(Concept)與寫作之間的關係，並以此作為基礎自動建立一套中文作文評分系統。本系統以概念為主要特徵(feature)，系統建立的過程共涉及三個模組：1.Generation-特徵產生 2.Selection-集合的篩選 3.Evaluation-特徵評估。 本系統以知網(Hownet)作為概念空間，並將作文的用詞對應到知網中予以概念化。其次，我們設計一個方法從概念空間中自動有效率的篩選出其中的子集合，並且設計了一個評估函數來計算該集合的鑑別力。根據實驗結果，本系統評閱的正確率可達92%~95%，可作為協助老師批閱作文時的工具。

This thesis explores the relationship between concepts of words and essay writings. Based on their relationship, an automated scoring system can be automatically constructed. The system consists of such three models as Generation model, Selection model, Evaluation model. The system first transforms essays into the concept space defined by Hownet. Next we develop an efficient method to select subset of concept space to be used for feature subset. Lastly, an evaluation function is designed to evaluate the effectiveness of the subset for grading essays. The experimental results show that our system has achieved 90~95% correctness, it can greatly facilitate human grader.