If we assume semantic features that do not have to be irreducible and that are allowed to take on continuous values, they could in principle provide all the functionality that distributional representations offer, if they were sufficiently finegrained – but one of the core points that we want to make is that it is too hard to determine such a fine-grained set of semantic primitives. In contrast, it is doable to collect a large number of features automatically that are not (and do not need to be) individually inference-enabling, just in the aggregate. These features can be textual or non-textual.
展开▼