A tech company has built a neural machine translation system that translates English text to Spanish. While the system quickly generates translations that capture the general meaning, users complain about frequent grammatical errors and unnatural-sounding sentences. Based on your understanding of common paradigms for sequence-to-sequence tasks, propose and describe a two-stage architectural modification that could address this issue without retraining the initial translation model from scratch.