A Mighty Dataset for Stress-Testing Question Answering Systems

摘要

The general goal of semantic question answering systems is to provide correct answers to natural language queries, given a number of structured datasets. The increasing broad deployment of question answering (QA) systems in everyday life requires a comparable and reliable rating of how well QA systems perform and how scalable they are. In order to achieve this, we developed a massive dataset of more than 2 million natural language questions and their SPARQL queries for the DBpedia dataset. We combined natural language processing and linked open data to automatically generate this large amount of valid question-query pairs. Our aim is to assist the benchmarking or scoring of QA systems in terms of answering questions in a range of languages, retrieving answers from heterogeneous sources or answering massive amounts of questions within a limited time. This dataset represents an ideal choice for stress-testing systems' scalability, speed and correctness. As such it has already been included into the Large-scale QA task of the Question Answering Over Linked Data (QALD) Challenge and the HOBBIT project Question Answering Benchmark.