Building and Benchmarking Large Language Models for Machine Translation in Social Network Services

The global proliferation of online social interaction has increased demand for machine translation (MT) on social networking services (SNS), but conventional methods struggle with culturally distinctive content- such as memes, slang, and pop culture references. Although large language models (LLMs) have advanced general- purpose translation, their performance on SNS- specific scenario remains limited due to a shortage of specialized training data and evaluation benchmarks. To address this, we present RedTrans, a family of LLMs at multiple scales tailored for SNS translation, trained on a large- scale corpus via two stages: (1) supervised fine- tuning with Dual- LLM Back- Translation Sampling, an unsupervised selection strategy that uses LLM- based back- translation to curate diverse data for fine- tuning;and (2) Rewritten Preference Optimization (RePO), an algorithm that detects and corrects erroneous preference pairs through expert annotation to construct reliable prefer- ence corpora. We further introduce RedTrans- Bench¹¹https://github.com/HC-Guo/RedTrans, the first benchmark for SNS translation, covering phenomena such as humor localization, emoji semantics, and meme adaptation. Extensive experiments show that RedTrans outperforms state- of- the- art LLMs and has been successfully deployed in real- world environments, demonstrating that its effectiveness in bridging the gap between generic and domain- specific translation.