Graph Traversal-based Question Answering System over DBpedia

• Main Authors: Jia-Hui Lai, 賴家慧
• Other Authors: 呂瑞麟
• Format: Others
• Language: zh-TW
• Published: 2018
• Online Access: http://ndltd.ncl.edu.tw/handle/2b42b3

碩士 === 國立中興大學 === 資訊管理學系所 === 106 === We present a natural language question answering system that queries DBpedia to convert user''s natural language question into a SPARQL structure queries over linked dataset. First, we identify useful entities from the question, then calculate the quantity and count the quantity to select an pivot from the question sentence, then re-create the tree with pivot and remove the stopwords to produce a graph. According to the sub-graph to traverse and query DBpedia sequentially, and then get one or more sets of triples. Then combine the triples with other additional conditions, finally generate the complete SPARQL syntax and get the answer. In this paper, we used QALD-7, QALD-6, QALD-5, QALD-4 and QALD-3 multilingual test dataset to evaluate our method. In the complete dataset, we achieve an average precision of 0.19, an average recall of 0.23 and an average F-measure of 0.20 on the QALD-7；an average precision of 0.31, an average recall of 0.54 and an average F-measure of 0.34 on the QALD-6；an average precision of 0.33, an average recall of 0.43 and an average F-measure of 0.36 on the QALD-5；an average precision of 0.25, an average recall of 0.34 and an average F-measure of 0.24 on the QALD-4；an average precision of 0.40, an average recall of 0.50 and an average F-measure of 0.41 on the QALD-3 and further in the query type question on QALD-3, we use an average F-measure to compare and achieve 0.32 on Aggregation, 0.26 on List and 0.48 on Other. Experimental results show that our method can solve complex questions.