A better entity detection of question for knowledge graph question answering through extracting position-based patterns

Abstract

Abstract Entity detection task on knowledge graph question answering systems has been studied well on simple questions. However, the task is still challenging on complex questions. It is due to a complex question is composed of more than one fact or triple. This paper proposes a method to detect entities and their position on triples mentioned in a question. Unlike existing approaches that only focus on detecting the entity name, our method can determine in which triple an entity is located. Furthermore, our approach can also define if an entity is a head or a tail of a triple mentioned in a question. We tested our approach to SimpleQuestions, LC-QuAD 2.0, and QALD series benchmarks. The experiment result demonstrates that our model outperforms the previous works on SimpleQuestions and QALD series datasets. 99.15% accuracy and 96.15% accuracy on average, respectively. Our model can also improve entity detection performance on LC-QuAD 2.0 with a merged dataset, namely, 97.4% accuracy. This paper also presents Wikidata QALD series version that is helpful for researchers to assess the knowledge graph question answering system they develop.

Keywords

• Entity detection
• Entity recognition
• Knowledge graph question answering
• Complex question
• Question pattern