A fundamental task in natural language processing is the efficient representation of lexica. From a computational viewpoint, lexica need to be represented in a way directly supporting fast access to entries, and minimizing space requirements. A standard method is to represent lexica as minimal deterministic (classical) finite-state automata. To reach such a representation it is of course possible to first build the trie of the lexicon and then to minimize this automaton afterwards. However, in general the intermediate trie is much larger than the resulting minimal automaton. Hence a much better strategy is to use a specialized algorithm to directly compute the minimal deterministic automaton in an incremental way. In this chapter we describe such a procedure.

In this chapter we explore deterministic finite-state transducers. Obviously, it only makes sense to ask for determinism if we restrict attention to transducers with a functional input-output behaviour. In this chapter we focus on transducers that are deterministic on the input tape (called sequential or subsquential transducers). We shall see that only a proper subset of all regular string functions can be represented by this kind of device and we describe a decision procedure for testing whether a functional transducer can be determinized. Further we present a subsequential transducer minimization procedure based on theMyhill–Nerode relation for string functions.