Title: AutoKG: Constructing Virtual Knowledge Graphs from Unstructured Documents for Question Answering

Abstract: Knowledge graphs (KGs) have the advantage of providing fine-grained detail for question-answering systems. Unfortunately, building a reliable KG is time-consuming and expensive as it requires human intervention. To overcome this issue, we propose a novel framework to automatically construct a KG from unstructured documents that does not require external alignment. We first extract surface-form knowledge tuples from unstructured documents and encode them with contextual information. Entities with similar context semantics are then linked through internal alignment to form a graph structure. This allows us to extract the desired information from multiple documents by traversing the generated KG without a manual process. We examine its performance in retrieval based QA systems by reformulating the WikiMovies and MetaQA datasets into a tuple-level retrieval task. The experimental results show that our method outperforms traditional retrieval methods by a large margin.