SEOUL – NOV. 15, 2021 –  Netmarble, a leading developer and publisher of high-quality mobile games, became the first Korean gaming company to rank first in an Automatic Post Editing (APE) Task, presented this year at EMNLP (Empirical Methods in Natural Language Processing 2021).

The EMNLP, held annually, is a world-renowned global conference on statistical machine translation. This year, it was held in Punta Cana, Dominican Republic and online from November 7th to 11th. Each year, EMNLP features shared tasks in various areas of machine translation including automatic post editing, news translation, biomedical translation, and more.

The success of Netmarble’s APE system comes from Curriculum Training Strategy (CTS) and Multi-task Learning (MTL) methods. CTS was applied across multiple training stages, allowing the system to learn in preparation for complex problems or target tasks in later stages. MTL alleviates the data sparsity problem by sharing knowledge representations among tasks and categorizing relevant data.

“The dataset of WMT21’s APE shared task was already high in quality and the dataset size is not big enough. To this end, we applied CTS with an additional large dataset provided by the competition side to bootstrap the performance, and another key point is we make use of MTL during the fine-tuning step which shows the best performance in human evaluation metric.” Said Shounan An, Team Manager, Netmarble AI Center. “The APE system developed by Netmarble displayed low Translation Error Rates and high BLEU(Bilingual Evaluation Understudy Score), resulting in a combined score that surpassed those of other global top platform companies.”

Oh Insoo, Director of Magellan Division, Netmarble AI Center stated, “As a global game publisher and developer, Netmarble already incorporates AI translation technology in a variety of games, but through the introduction of an AI system in post-editing, we were able to optimize both translation quality and time.” And added, “By focusing on CTS and MTL, we plan to introduce a system capable of accurate translation without the need for human intervention.”