Beyond Corpus Lookup : Towards Heuristic Reasoning with Text

Both knowledge-based and text-based approaches to question answering suffer from brittleness. Textbased approaches use existing corpora like the web to answer a broad array of questions. However, most of these systems rely on corpus lookup. They have limited abilities to combine information from multiple sources, draw inferences, double check and explain answers. Knowledge-based systems can make sophisticated inferences, however, currently there exists no knowledge base that is broad enough to perform reasonably in the TREC question answering competitions. We argue that both these approaches can leverage off each other at heuristic reasoning, which is the type of reasoning underlying the human ability to make educated guesses. We focus on back of the envelope reasoning, the process of generating ballpark quantitative estimates. Starting from our previous work on a knowledge-based back of the envelope problem solver [Paritosh and Forbus, 2005], we present an analysis of how back of the envelope reasoning can be done with text.