This is an interesting short paper from ACL 2011. The paper examines the problem of “recognizing textual entailment,” essentially that of judging whether or not a conclusion is justified from a given piece of text. That’s interesting because drawing conclusions from text depends on a large amount of assumed knowledge in the reader, and machines don’t generally have that knowledge. This paper examines what kinds of knowledge are typically required for making these kinds of judgments.

Why is it important for any practical application? Well, if you are building a system that wants to learn through reading a bunch of text, and perhaps be able to answer questions about what you’ve read, you will do better if you can get more information (and more accurate information) out of that text. And if you have these kinds of reasoning capabilities, you will make better conclusions about what you are reading. It was an interesting piece of work.