... achieved at the end of Model 1 training, and at 5 iterations of HMM. When random initializationis used, we run 20 random trials with different ini-tialization, and report the min, max, and mean ... performsworse for a very small data size, but it catches up and surpasses the random models at data sizes greaterthan 100 sentence pairs.To further evaluate the impact of initialization forIBM Model 1, ... different random initial-ization in EM, and the impact of initialization on testset log-likelihood and alignment error rate. Theseexperiments suggest that initialization does matterin practice,...