Went to an interesting talk today about multi-hop reasoning on knowledge graphs using surface language. (E.g. how do you reason using the language that is on the page or in the air when the reasoning involved requires multiple inferences). The gist was that it doesn't work well if the error rate of mapping onto the meanings is much above 0 (one tenth of one percent was too much error for three hops). Pretty cool stuff to think about for creating implicature mining systems.