Title: Causal Knowledge Graphs for Natural Language Understanding
Speaker: Aditya Kalyanpur
Abstract
At Elemental Cognition (EC), we are building AI systems that can read, reason and understand text by building logical and causal models that capture its underlying meaning. A fundamental challenge is that AI systems typically lack common-sense background knowledge needed to build rich text-interpretation models. To address this issue, two problems need to be solved: first, we need to acquire often-implicit highly varied commonsense knowledge at scale; and second, we need to develop efficient techniques to incorporate this knowledge into an AI reasoning system. At EC, we have developed solutions to both these problems - we have created GLUCOSE, a high-quality semi-structured commonsense knowledge graph of ~600K causal rules that was crowd-sourced (using Amazon Mechanical Turk), and then we use GLUCOSE knowledge as seed data to fine-tune large pre-trained language models to construct "dynamic rule generators" that are plugged into our neuro-symbolic reasoner, Braid. In this talk, I will describe how we built and use GLUCOSE to address the knowledge acquisition bottleneck for an AI reasoning system that does language understanding.
Slides
Bio
Adi Kalyanpur is a senior research scientist at Elemental Cognition where his focus is on integrating symbolic and statistical approaches to solving the language understanding problem. His work deals with various aspects of text understanding, including efficient reasoning algorithms combining formal inference with statistical models, and scalable knowledge acquisition methods (automatically from corpora, or from end-users/crowdsourcing). Prior to joining Elemental Cognition, he was a research staff member at the IBM T.J. Watson Research Center, where his work spanned the areas of knowledge representation, automated reasoning, natural language processing, and machine learning. Adi was one of the key technical leads on the IBM Watson Question Answering system that won the Jeopardy! challenge. Before Watson, he helped build SHER, a highly scalable logical reasoning technology that was deployed in IBM products. Adi completed his Ph.D. in Computer Science from the University of Maryland at College Park.