Abstract

Abstract: Knowledge Graphs (KGs) are graph-structured Knowledge Bases where facts are encoded by links between entities, and their use is pervasive both in industry (e.g. see Google Knowledge Graph, Microsoft Satori) and academia (YAGO, DBpedia, NELL , OpenCyc). KGs are extremely useful, as they provide automated systems with machine-readable domain specific knowledge, allowing AI systems to perform deductive (and inductive) reasoning in a variety of domains. The talk will discuss about Representation Learning in KGs, its shortcomings, how we can introduce First-Order Logic constraints in the learned representations, and how we can extend such techniques to NLP models. Then the speaker will talk about his recent work in neuro-symbolic reasoning – namely on Neural Theorem Provers [1, 2] – and how we can perform explainable deductive reasoning jointly on Knowledge Graphs and Natural Language at scale.

[1] https://arxiv.org/abs/1705.11040

[2] https://arxiv.org/abs/1807.08204

Bio: Pasquale Minervini is a Postdoctoral Research Associate in Statistical Natural Language Processing and Machine Learning in the UCL Machine Reading group. He is funded by a Machine Reading grant from the Allen Institute for Artificial Intelligence (AI2).