小學生問答系統之研究：以歷史地理領域素材為例

Abstract

本文提出一個處理小學五、六年級學生問題的問答系統，問答系統以半自動化方式從選擇題中建立知識本體，藉以分析問題的類型，並以小學五、六年級的教師手冊做為知識來源，使用solr建立資訊檢索系統，找出教師手冊中的段落，最後找出可能的答案並回傳。 本研究將問答系統分成數個模組，實驗時分別觀察在康軒選擇題資料集及以及小學生問句資料集中的表現。在每個模組中，分別考慮不同方法，並且比較其效能，同時分析不同方法之間的優缺點。 實驗結果顯示，問答系統可以對小學生問句所帶有的問題進行處理，並正確的找出問題中有用的資訊。問答系統在康軒選擇題資料集中，平均倒數排名(mean reciprocal rank)達到0.52，在小學生問句資料集中則達到0.46。

In this work, we present a question-answering (QA) system for the 5th and 6th grade students to support their learning. Students ask questions and the QA system gives answers to their questions based on a knowledge base created from textbooks. A taxonomy is constructed semi-automatically from multiple choice questions (MCQs) and student questions. It is used to find the probable answer type of a question. Then, we construct an IR system using Solr, an open source IR system, based on textbook knowledge. We use textbook paragraph returned by solr to determine answers. In the experiments, we examine each component of the question-answering system and analyze the possible errors separately. We consider different approaches and show the pros and cons of each approach. The experimental results show that our question answering system can handle the questions submitted by children. Our system achieves 0.52 mrr score in textbook question dataset and 0.46 in students’ question dataset.