在數位學習導入答題系統通常可作為虛擬助教，減輕老師負擔，隨時為學生提供幫助。給定文章段落及問句，多選項之閱讀理解任務需要模型從選項集預測正確答案，除了使用強大的語言模型作為編碼器，此任務常常需要透過文章、問句及選項的訊息比對，取得三者之間的關聯性，且目前許多方法只考慮問題感知之文章表示法而忽略文章感知之問題表示法。 本論文提出一個選擇題答題系統，利用中學課文做為知識來源，將文章、問句及選項序列透過編碼器轉換成含上下文訊息之向量表示法。我們模擬人類作答的思維，加入兩個解題方法：(1)文章句子選擇，擷取與問句關聯性最高的文章句子，(2)選項交互參考，四選項互相比對訊息後進行編碼，在我們的題庫上有更好的效果。再連接雙向多頭注意力機制產生進階表示法，經分類器產生預測答案。對公開資料集及我們的歷史題庫的實驗結果顯示，我們的模型與基準模型相比有更好的效能。 ;Given a passage and a question, multi-choice reading comprehension tasks require model to predict the correct answer from a set of candidate answer options by using a strong language model as encoder. The tasks usually need information comparison between passage, question and answer option to get the relevance among them. Most of the existing methods consider the question-aware passage representation but not passage-aware question representation. In this paper, we will present the ‘PADMA’ stands for Passage sentence selection and Answer option interaction integrated on Dual Multi-head co-Attention. We propose a multi-choice question answering system PADMA, which collects the middle school textbook as a knowledge source, and encodes the sequence formed from passage, question and answer option to a contextualized vector representations. We simulate the way human solve multi-choice problem, and integrate two reading strategies: (1) passage sentence selection, which helps decide the most relevant sentence from the passage corresponding to the question; (2) Answer option interaction, which encodes bilinear representations between each two options. Then a dual multi-head co-attention model is used to generate the advanced representation and a decoder is to calculate the answer prediction. The experiment result shows that our proposed model achieves a better performance compared with base models.
