利用廣義知網與維基百科產生單一選擇題之學習式架構

Abstract

自動出題系統是建立e-Learning學習環境的一種重要工具，也教師減少出題考試的負擔。目前已經有許多文獻在探討自動出題的可行性，現有的文獻指出，自動出題系統的確可以減輕教師們出題的時間，並且在某些特殊的題型，例如英文克漏字測驗，已經可以達到作為課堂考試用的材料的程度。在本論文中，我們將利用知識本體的資料庫來建立一個能夠跨領域的自動出題系統，並且利用維基百科的條文補足知識本體資料庫本身的不足，來克服人工產生的知識本體資料庫本身的缺陷。本論文著重在單一選擇題的出題設計，利用機器學習的方法來判斷怎樣的句子包含事實的知識，找出適合出題的句子，同樣再利用機器學習的方法來找出適合挖空成為考題的詞，並且利用知識本體資料庫裡的語義定義以及知識結構，來取出意思相近的干擾項，若是我們要產生的干擾項的字詞並沒有預先被定義在知識本體的資料庫裡，將會查詢維基百科是否有該字詞的描述條文，若有則取回，同時取回所有在知識本體裡的字詞的維基百科描述，將相同字詞利用非監督式的分群方法歸類該字詞，再產生與其語意上相近之干擾項。因為每一個題目只能有一個正確答案，所以接下來我們要確認其他的干擾項是否也可以是正確答案，若是的話，我們要丟棄該干擾項並找其他的選項來取代。我們是利用Google 搜尋的結果數量當作檢查標準，當尋的結果數量超過某一個閥值或是比正確答案的選項高或在同個數量級時，便視該干擾項很可能也是正確答案，便會將其刪除。最後，我們將展示系統給做一般使用者測試，搜集實際答題紀錄，並做系統滿意度調查。在使用者測試的結果顯示，我們系統達到70.4%的可接受率。

Automatic question generation is an important tool for e-Learning environment. It is also a useful tool for teachers to reduce workload of generating questions. There are some related works discussed the possibility of automatic question generation. From existing literature, it shows that an automatic question generation system can save lots of effort and time for teachers. Moreover, in some special type of test such as cloze test, it can output materials that can be used in real classes. In this paper, we use E-HowNet ontology database to construct a cross-domain automatic question generation system. In order to overcome the coverage limitation of ontology database, we retrieve description from Wikipedia to cover the missing words of E-HowNet ontology. This paper focuses on generating multiple-choice questions. We use machine learning based methods to decide which kinds of sentences contain factual knowledge. After identifying the suitable sentences, we apply machine learning methods again to identify which word phrases are suitable to become blank parts. Then, we use the semantic definitions of ontology database to choose close but not the same meaning words to become distractors. If we cannot find the words of blank parts, then we try to retrieve their description and also other existing ontology words from Wikipedia. By applying an unsupervised clustering method, we can discover the relationship between the missing words and those existing words. Therefore, we can use the relationship and combine the missing words with existing ontology database to help us retrieve suitable distractors. Because each question have only one correct answer, we need to verify whether other distractors could be the correct answer or not. If yes, then we should discard it and find another one. We retrieve the number of Google search results of the sentence which fills the correct answer back to the original sentence. Then, compare it with other distractors’ Google search results. If the number of results of a distractor exceeds a certain threshold or has the approximate number results to the number of the correct answer, we will assume it has a higher possibility that also could be a correct answer. If that is the case, we will discard it and choose another one. Finally, we conduct a user study to analyze the usability of our results. We invite some people to evaluate our system. The evaluation result indicates that our prototype system is feasible to generate multiple choice questions in 70.4% acceptable rate.