Natural Language Processing: Structure and Complexity

We introduce a method for analyzing the complexity of natural language processing tasks, and for predicting the difficulty new NLP tasks. Our complexity measures are derived from the Kolmogorov complexity of a class of automata --- {\it meaning automata}, whose purpose is to extract relevant pieces of information from sentences. Natural language semantics is defined only relative to the set of questions an automaton can answer. The paper shows examples of complexity estimates for various NLP programs and tasks, and some recipes for complexity management. It positions natural language processing as a subdomain of software engineering, and lays down its formal foundation.