Guesswork for Inference in Machine Translation with Seq2seq Model
Litian Liu, Derya Malak, Muriel Médard · 2019
One-shot inference is used in machine translation today. In practice, the output probability distribution is not concentrated since there might be multiple valid translations. Therefore, we propose to use a multi-shot inference mechanism in this paper. We analyze the Markovian property of sequence to sequence (seq2seq) model. Based on a large deviation principle satisfied by guesswork on Markov process, we derive theoretical upper bounds on the accuracy of the seq2seq model with single correct answer under one-shot inference and multi-shot inference. We establish analogous bounds when there are multiple correct answers in translating. We also discuss the extension of the results to translation with distortion tolerance.