From natural language processing to neural databases

In recent years, neural networks have shown impressive performance gains on long-standing AI problems, such as answering queries from text and machine translation. These advances raise the question of whether nets can be used at core query processing to derive answers facts, even when facts are expressed in natural language. If so, it is conceivable that we could relax fundamental assumption database management, namely, our data represented fields a pre-defined schema. Furthermore, technology would enable combining information text, images, structured seamlessly. This paper introduces databases , class systems use NLP transformers localized answer derivation engines. We ground vision NeuralDB, system for querying short language sentences. demonstrate models, specifically transformers, select-project-join if they given set relevant facts. However, cannot scale non-trivial nor set-based aggregation queries. Based these insights, identify specific research challenges needed build databases. Some require drawing upon rich literature others pose new opportunities community. Finally, show with preliminary solutions, NeuralDB already over thousands sentences very high accuracy.