Knowledge Induction Team @ IBM Research AI - Talks
Minimally Supervised Knowledge Graph Induction from Text
Inducing Knowledge Graphs (KGs) from enterprise data is a labor-intensive process requiring a collaboration between Subject Matter Experts (SMEs), Data Scientists and Knowledge Engineers. The goal of the Knowledge Induction department within IBM Research AI is to develop technology enabling our customers to build such assets with a focus on solutions requiring minimal domain adaptation effort from SMEs. In this talk, I’ll provide an overview of our research program. Specifically, I’ll present how to use deep learning architectures to induce types and relations from text using distant supervision, transfer learning, knowledge base completion and validation. I’ll show applications on financial data and virtual assistants.
Knowledge Graph Induction using Distantly Supervised Deep Nets
Knowledge Graph Induction using Distantly Supervised Deep NetsInformation Extraction (IE) analytics are needed to induce Knowledge Graphs (KGs) from text. Adapting IE to new domains is a labor-intensive process requiring Subject Matter Experts (SMEs), Data Scientists and Knowledge Engineers. In addition, a severe limitation of most IE systems is that they are typically unable to identify implicit relations between entities, severely limiting the applicability of KG induction technology to enterprise solutions. In this talk I will present the results of our long-term research program aimed at developing SME friendly technology to induce KGs from domain specific corpora with minimal effort. I’ll describe how to train deep nets for relation extraction using distant supervision and transfer learning by analogy from Linked Open Data and how to identify implicit relations in texts using unary relations, composite contexts and Knowledge Base Completion. Finally, I’ll present a knowledge induction framework able to induce taxonomies directly from text provided as input in any domain, enabling SMEs to easily customize them to their needs.
Knowledge and Reasoning in Cognitive Computing
IBM Watson is arguably one of the most advanced form of AI in the cognitive computing marketplace, as demonstrated by the historic exhibition match on the television quiz show Jeopardy!. Since then, cognitive technology is becoming pervasive in information systems. A wide range of cognitive capabilities, ranging from machine translation to image recognition have been made available on cloud to any developer, opening a new market for cognitive computing. However, perception is just one of the many fundamental abilities that characterize cognition. The new challenge for IBM Watson and for the whole industry will be to help human professionals doing better decisions in their own areas of interest. The next generation AI systems will show professional level of competence, leveraging deep domain knowledge accumulated over decades. To this aim, deep learning based solutions will be just one of many ingredients, knowledge representation and reasoning being the other two pillars. In this talk I’ll describe progress made by my team at IBM Research and envision new research directions.
Knowledge and Reasoning in IBM Watson