Neural Database Operator Model

Abstract

Our goal is to answer queries over facts stored in a text memory. The key challenge in NeuralDBs(Thorne et al., 2020), compared to open-book NLP such as question answering (Rajpurkar et al., 2016,inter alia), is that possibly thousands of facts must be aggregated to provide a single answer, without direct supervision. The challenges represented in NeuralDBs are important for both the NLP and database communities alike: discrete reasoning over text (Dua et al., 2019), retriever-based QA(Dunn et al., 2017) and multi-hop QA (Welbl et al.,2018; Yang et al., 2018) are common components.