Markov Chains for Robust Graph-Based Commonsense Information Extraction

Commonsense knowledge is useful for making Web search, local search, and mobile assistance behave in a way that the user perceives as “smart”. Most machine-readable knowledge bases, however, lack basic commonsense facts about the world, e.g. the property of ice cream being cold. This paper proposes a graph-based Markov chain approach to extract common-sense knowledge from Web-scale language models or other sources. Unlike previous work on information extraction where the graph representation of factual knowledge is rather sparse, our Markov chain approach is geared towards the challenging nature of commonsense knowledge when determining the accuracy of candidate facts. The experiments show that our method results in more accurate and robust extractions. Based on our method, we develop an online system that provides commonsense property lookup for an object in real time.