基於貝氏機器學習法之中文自動作文評分系統

Abstract

在本論文中，我們探討文章直接與間接特徵對於寫作評分之間的關係，並以此作為基礎，建立一套以貝氏機器學習法為主的中文作文自動評閱系統。在本研究中我們認為，文章的間接特徵(外在特徵)雖然無法提供足夠的語義資訊，卻往往深切的影響評分老師對於文章好壞評斷的第一印象；如文章字數、分段數、標點符號的正確使用與否等其他多項外在因素，皆為評分老師在尚未深入細讀文章內容時用以作為評分標準的圭臬。但一篇文章的好壞，不僅僅只是以外觀的特徵來決定，且須更進一步探討文章的各段內容。據此想法，本系統對於文章的評分流程共分為三個階段：1.Holistic Scoring-整體評鑑 2.Paragraphic Scoring-分段評鑑 3.Integration-評鑑整合。而根據實驗結果，本系統評閱的正確率可達95%~97%，是作為閱卷老師評分時的良好工具之一。

This paper proposes an efficient method based on Bayesian Theorem to score Chinese essay according to the direct and indirect features of an essay. It includes words, nouns, themes, oral writing, average number of words of a section and concepts of an essay. In this study, we determined the holistic and paragraphic score of a testing data firstly. Subsequently calculate the grade of the testing data by integrating the relation of holistic and paragraphic score. Experimental results show that our approach compares favorably with some other Automatic Chinese Essay Scoring (ACES) systems.