整合主體、類別和屬性識別的知識庫簡單問題問答系統

Abstract

隨著網際網路高度普及化，每天都有許多新知識產生。這些新知識經過整理後，以知識庫的形式儲存，如Freebase和DBpedia。有了這些豐富的資源，如何有效率地從中獲取需要的資訊是個很重要的課題。自然語言問答系統是最直接且貼近人們生活的一項應用，使用者可以用熟悉的語言提出任何問題，並透過問答系統從知識庫中獲取答案。 本研究提出一套識別知識庫中主體、類別和屬性的仿真陳述問答系統，以回答簡單類型的問題。我們首先提出數種新的特徵，使得問題中候選主體的排序更加準確。同時，我們也將知識庫中的關係分為類別和屬性，並分別以一個雙向長短期記憶模型進行識別。實驗結果顯示，我們的系統在SimpleQuestions資料集上，達到目前最好的效能。

With the popularity of the Internet, more and more new information is generated every day. The information can be stored in knowledge base, such as Freebase and DBpedia. To access the knowledge efficiently and quickly to acquire what users need, the most direct and close to people's life is question answering system in natural language. People can ask any questions in their familiar languages, and then use the question answering system to get answers from the knowledge base. This study presents an approach to identify subject, type and property from knowledge base for answering factoid simple questions. We propose new features to rank entity candidates in knowledge base. Besides, we split a relation in knowledge base into type and property. Each of them is modeled by a bi-directional long short-term memory for identification. Experimental results show that our model achieves the state-of-the-art performance on the SimpleQuestions dataset.