AsianFin -- NetEase Youdao has developed the 14B small-parameter domain-specific model based on its self-developed Ziyue Translation Large Model 2.0, a significant iteration of its translation large model, AsianFin has learned.

This new model features high-level translation performance while reducing computational resource consumption, lowering deployment costs, and enabling easier integration into existing systems and devices. The upgrade makes the technology more accessible for a wider range of practical applications.  

The new translation technology, powered by the large model, has already been implemented in Youdao Dictionary, Youdao Translation, and Youdao Translator. Users can now choose between two parameter options: a standard model and an advanced model, with seamless switching between the two.

Meanwhile, the large model has been integrated into NetEase Youdao's smart hardware products. The Youdao Dictionary Pen X7 series has already been upgraded to the latest model, with other devices set to follow.  

The performance of large language models is not solely determined by the number of parameters but also depends on data quality, domain adaptability, and algorithm optimization. NetEase Youdao's 14B small-parameter domain-specific model leverages advancements in data processing, utilizing high-quality translation corpus data meticulously annotated by certified English language specialists and professional translators. This extensive data repository enhances the model's ability to handle diverse translation scenarios effectively.  

On the algorithmic front, Youdao built upon the Ziyue large model and conducted secondary pre-training to create a translation foundational model that balances professional accuracy and domain specificity. Techniques such as large model distillation, model fusion, and Online DPO (Direct Preference Optimization) were employed to avoid catastrophic forgetting issues while significantly improving translation performance in terms of operational efficiency, accuracy, and fluency.  

To evaluate the model's performance, Youdao developed a translation assessment tool called the Reward Model, which utilizes accumulated translation data to provide a reliable quantitative basis for evaluation. This is complemented by a comprehensive manual evaluation framework, enabling multi-dimensional analysis of the model's translation results.  

The Ziyue Translation Large Model 2.0 marks an improvement in Chinese-English translation, particularly in vertical scenarios. Internal evaluations by Youdao indicate that the new model demonstrates higher accuracy and fluency across 19 vertical domains, including humanities, business, lifestyle services, healthcare, and science. It outperforms previous versions in professionalism, accuracy, linguistic conventions, and style.  

A person in charge of NetEase Youdao emphasized the importance of vertical models, saying, "General large models compete on parameters and computational power, but translation cannot achieve professionalism solely through parameter stacking. While general large models race to scale up, we firmly believe in the future value of vertical models. Addressing pain points in professional scenarios with specialized applications is what we are doing."  

Prior to the rise of the large model technology, Youdao's translation solutions were primarily based on statistical machine translation and neural machine translation (NMT). Today, Youdao's translation products boast over 1 billion users. According to Quest Mobile, the NetEase Youdao Dictionary has surpassed 100 million monthly active users and has consistently ranked first in the education tool category for six consecutive years since 2019.  

更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App