Question Answering Using Semantic Query Graphs: A Replication Study

Abstract

Semantic methods for natural language question answering are better suited for domain-specific applications than their deep-learning counterparts because they can utilize domain-specific structured data and do not require the large training datasets typically necessary for deep learning. Zou et al. describe a method for RDF-based question answering in which the input question is transformed into a semantic query graph that is subsequently mapped to a subgraph of the domain knowledge graph containing the answer. In this work we replicate this approach and evaluate it using the Question Answering over Linked Data (QALD) dataset. Of the 131 relevant questions in this dataset, the method answers 41 questions, 23 of which are answered correctly, achieving an Fl-score of 0.201. Integration of state-of-the-art named entity recognition (NER) techniques provide a path forward for improving performance.