Analytics

XL8 has made a significant contribution to the field of natural language processing by submitting a noteworthy paper to the 2023 Conference on Empirical Methods in Natural Language Processing, one of the world’s premier conferences for NLP.

Their paper, titled “Enhanced Simultaneous Machine Translation with Word-level Policies,” authored by Research Scientists Kai Kang Kim and Matthew Hankyu Cho, both of whom contributed equally to the work, was accepted for publication in the upcoming EMNLP 2023 Findings in December.

This research paper introduces an innovative method aimed at enhancing the performance of Simultaneous Machine Translation (SiMT). The key innovation presented is the incorporation of word-level policies into real-time translation systems, departing from the traditional practice of using subword-level policies.

Typically, Machine Translation models segment words into subwords for various advantages like flexibility in modeling and handling rare words. For SiMT models, which operate under a policy that determines when to receive remaining input words or resume translation, previous approaches relied on policies that worked at the subword level.

However, a subword-level policy often produces translations from partially processed input words before all of their subwords are processed, resulting in suboptimal translation quality.

Moreover, it hinders the integration of a language model (LM), which has been proven to be a powerful method for enhancing many NLP tasks, due to the subword mismatch between the LM and SiMT.

This study demonstrates that by adopting the proposed word-level policies, it is possible to address these limitations and achieve more accurate results with lower latency, hence significantly improving real-time translation performance.

The real-time translation model developed through this research has already been seamlessly integrated into EventCAT, an advanced real-time translation and interpretation service offered by XL8.

With XL8’s EventCAT, you can elevate the accessibility and inclusivity of your live broadcasts and event streams while providing a superior user experience.

It also generates on-screen translated text and synthetic speech, enabling multilingual conversations for global call centers, conferences, international meetings, or video calls.

Notably, EventCAT has played a pivotal role in offering mobile translation services in multiple languages, such as English, Arabic, and Chinese, at various domestic and international events like Nature Forum, Leeum Art Museum’s Special Event, ComeUp2023, and Content Universe Korea.

This integration has transformed the event experience, enabling participants to fully immerse themselves in the content, regardless of language barriers.