... for bothtraining and test pairs. We do it in two stages: ex-tracting similar nominals, and obtaining more con-texts.4.1.1 Extracting more wordsFor each nominal pair (w1, w2) in a given ... is donewithout accessing any training data, anticipating thatmost meaningful relationships, including those in agiven classification problem, will be represented bysome of the discovered clusters. ... CW1 [In x] CW2[Postfix]where Prefix, In x and Postfix contain only HFWs.We require Prefix and Postfix to be a single HFW,while In x can contain any number of HFWs (limit-ing pattern length by window...