Bidirectional Modeling for Simultaneous Neural Machine Translation

Anonymous

Abstract: Simultaneous Neural Machine Translation (SimulNMT) generates the output before the entire input sentence is available and only uses the unidirectional attention from left-to-right so that its decoding highly relies on future forecast according to word ordering rules. However, it is utopian that the word order strictly obeys the grammar rules in a language, especially in oral. To address the mismatch between SimulNMT expecting strict word order and free word order in real scenario, we propose a bidirectional modeling. In detail, we train another backward model where the input sentence is from right-to-left and keep the target sentence from left-to-right. Then we join this backward model into the standard forward SimulNMT model during decoding. This strategy enhances the robustness of SimulNMT and empowers the model to be more adaptable for the inconstant word ordering phenomenon. Experiments show that our method brings improvement over the strong baselines.