On the Complexities of NLP Systems

We give an account of the semantic differences between NLP systems, e.g. those used for information extraction and the older ones such as BORIS or ELIZA. We start by defining the concept of semantic complexity, and then compare NLP systems wrt. to different complexity measures. We argue that BORIS could not control the complexity of interaction, and that is why it did not scale up: its iterated why-complexity was orders of magnitude too large. In contrast, by limiting the extracted information to simple semantic types, current systems limit their semantic complexity.