Simulation of Fixed Length Word String Probability Distributions

I. WORD STRING DISTRIBUTION SIMULATION The aim of this description is to provide a framework for simulating specific word string distributions whose general structure follows the structure of word string posterior distributions relevant for string recognition. For the specific case of string classes w 1 = w1, w2, . . . , wN of fixed length N with words wn ∈ V , with vocabulary size V = |V|, and corresponding observation vectors per word position, x1 = x1, . . . , xN , we consider simulations of three types of distributions to obtain word string distributions: 1) global context dependence: simulate p(w 1 |x1 ) directly by a normalized distribution over V N (string) classes; 2) independent emission probabilities and bigram prior: simulate p(w|x) and p(w|v) to obtain p(w 1 |x1 ) = N ∏