For many semantic analysis tasks, including relation analysis, it would be desirable to have a description of words (or rather, the concepts behind them) in terms of semantic primitives. Such representations can be built only for small sets of words/concepts. When we deal with unrestricted vocabulary, there are various resources that we can turn to, to extract features to replace semantic primitives.

We will see several sets of features extracted from different sources, and how they compare for semantic relation analysis. We will see that no set gives the best performance for all relations. Looking at which description works best for which relation, will give interesting information about the relations themselves.