A Multi-Time Selection Framework for Machine Translation Based on Large Language Models
Xiaolei Dong · 2025
Machine translation, as a key technology for eliminating language barriers, plays an irreplaceable role in the process of globalization. In recent years, large language models have significantly improved the quality of cross-language text generation by virtue of their powerful semantic understanding ability. However, the random generation characteristics of such models lead to large fluctuations in the single inference results, making it difficult to present their optimal performance stably. In this paper, we propose a multi-time selection framework (MSF) for machine translation based on large language models, which leverages multiple generation and reverse validation, which improves the translation stability of large language models through zero training cost. The method first drives the large language model to perform multiple rounds of parallel translation of the source text, and then builds a two-way evaluation mechanism: calculating the semantic fidelity of the candidate translations in the forward dimension, verifying the generation quality through back-translation consistency in the reverse dimension, and ultimately dynamically selecting the optimal translations based on the confidence maximization criterion. The experiment constructs a test environment of 8 groups of mainstream language pairs based on the WMT benchmark dataset, and the results show that the method improves 1.763 points on average in BLEURT compared with single generation, which significantly alleviates the problem of random fluctuation in the generation results. This study provides a reproducible technical path for exploiting the translation potential of large language models.