Faster Machine Translation Ensembling with Reinforcement Learning and Competitive Correction
This work addresses computational bottlenecks in machine translation ensembling for researchers and practitioners, though it is incremental as it builds on existing CSB and FB frameworks.
The paper tackles the computational inefficiency of ensembling neural machine translation models by introducing SmartGen, a reinforcement learning-based strategy that selects a small, fixed number of candidates and uses feedback for training, along with a Competitive Correction Block, resulting in improved translation quality with reduced overhead.
Ensembling neural machine translation (NMT) models to produce higher-quality translations than the $L$ individual models has been extensively studied. Recent methods typically employ a candidate selection block (CSB) and an encoder-decoder fusion block (FB), requiring inference across \textit{all} candidate models, leading to significant computational overhead, generally $Ω(L)$. This paper introduces \textbf{SmartGen}, a reinforcement learning (RL)-based strategy that improves the CSB by selecting a small, fixed number of candidates and identifying optimal groups to pass to the fusion block for each input sentence. Furthermore, previously, the CSB and FB were trained independently, leading to suboptimal NMT performance. Our DQN-based \textbf{SmartGen} addresses this by using feedback from the FB block as a reward during training. We also resolve a key issue in earlier methods, where candidates were passed to the FB without modification, by introducing a Competitive Correction Block (CCB). Finally, we validate our approach with extensive experiments on English-Hindi translation tasks in both directions.