Plamo AI Translation Generates Human-Like Japanese

The service was introduced by Imajo, a researcher at AI startup NE Networks and a former Google engineer, who also helped launch the competitive programming site AtCoder during his student days. At a recent interview, Imajo demonstrated how Plamo can translate both ways between Japanese and English and even render entire web pages into natural Japanese, producing text that reads as if it were natively authored.

According to Imajo, existing translation models often struggle with long sentences and consistency in phrasing, but Plamo was trained extensively on longer texts, enabling it to capture fluency and flow. The system can also handle idiomatic expressions, such as correctly interpreting "a household on fire" as "financially strapped," avoiding literal missteps.

Plamo is built on a Japanese-developed large language model called Plamo 2, which was later customized into Plamo 2 Trans specifically for translation. Unlike conventional models trained primarily on parallel bilingual corpora, Plamo’s base model was pretrained separately on vast amounts of Japanese and English, allowing it to better understand word meaning and context before fine-tuning for translation.

A distinctive feature of the system lies in its tokenizer, which segments Japanese and English text differently from other models. For example, it groups words together in units that capture natural language flow, a method not commonly adopted by overseas models. Imajo explained that this approach helps maintain consistency in translation and contributes to the smooth readability of the output.

The service is currently available free of charge through browsers until the end of September. While large U.S. models dominate the global AI landscape, Imajo emphasized the importance of developing compact domestic models optimized for Japanese, noting that such models can achieve superior results in areas where local linguistic and cultural knowledge is critical.

Source: テレ東BIZ