An official from Flitto (CEO Jeong-su Lee) said, “Recently, the corporate’s fundamental task shouldn’t be technology development, but event monitoring.” Because of this the variety of places searching for this company’s real-time interpretation and translation solutions has increased significantly.
In reality, over the past few months, the answer has been significantly expanded to incorporate Incheon Airport, Seoul World Cup Stadium, a 60-passenger restaurant in Jongno-gu, Seoul, and the Busan Station ticket window. The answer was also implemented in Olive Young’s global mall. We participated in international events almost every month, including the Middle East’s largest IT event ‘GITEX’, Japan’s largest energy event ‘Japan Energy Summit’, and the UAE ‘International Power Forum’.
Nevertheless, he explained, “Participation in these events is ultimately a technology development process.” Because of this ‘living data’ that may improve the performance of artificial intelligence (AI) interpretation and translation solutions may be collected.
Flitto announced on the 18th that it shouldn’t be only proving its technological prowess through support for offline events, which have rapidly increased, but can be strengthening its data and models. In other words, the reason is that a ‘virtuous cycle’ of increased data collection and improved model performance was achieved.
Translation is a tough field where errors usually tend to occur than you may think. The pronunciation of the identical word varies depending on the speaker, and the identical word often has a distinct meaning. In other words, regardless of how good a model is, it may well be useless whether it is difficult to deal with various situations that occur in the sphere.
An example was given on the Japan Energy Summit. One participant explained that when he spoke English with a robust Tamil accent from southern India, a situation arose where even Americans had no selection but to give attention to a translation solution.
The reason is that Flitto’s ‘Live Translation’ can solve this problem. This can be a translation engine that goes beyond easy machine translation and combines STT (speech-to-text) and TTS (text-to-speech). Capture key keywords and conduct cross-contextual inference. Consequently, even when there are pronunciation differences or some uncertainties, accurate translation is feasible in line with the context. It’s a structure that ‘even if you happen to speak like a chunk of cake, it is known like a chunk of cake.’
The indisputable fact that it may well deal with such diverse accents and dialects is the most important reason why Flitto’s solution has received positive reviews one after one other not only in the US and Japan, but additionally within the Middle East and Southeast Asia.
One other strength of Flitto is ‘proper noun recognition’. Recognizing homonyms is basic. It’s explained that even essentially the most difficult ‘multilingual mixed use’ recognition is feasible.
An example was given of a Japanese person traveling to Korea. When a Japanese tourist asks the best way to get to City Hall Station, several situations can arise. The word City Hall Station may be uttered in various ways, akin to ‘City Hall Eki’, ‘City Hall York’, ‘City Hall Station’, or ‘City Hall York’. Korean and Japanese may be mixed, and as a consequence of the character of the Japanese language, the identical Chinese character can have multiple pronunciations. The Flitto solution can recognize all of those cases with the identical word.
There are also cases where the pronunciation difference exceeds. In Chinese, the word ‘Yun Yan’ (云烟) means cloud, but it is usually an abbreviation for a brand of cigarettes sold in Yunnan. In these cases, proper translation is unimaginable unless you will have actual data.
Each time an event ends, Flitto immediately collects logs for every solution, analyzes them, and learns them back into the engine with content appropriate for the context. Due to this fact, as event participation increases and the scope of application expands, the accuracy of translation improves to an extent that can’t be in comparison with other places.
Amongst these, generative AI is a key element that strengthens Flitto’s strengths of ‘understanding context’ and ‘understanding latest words and abbreviations’.
Translation using generative AI takes into consideration your complete context and linguistic characteristics and provides translation results that reflect the style and tone appropriate for the meaning and purpose of the text. Specifically, the consistency of long-text translation is a notable advantage.
After all, there are a lot of places that provide translation services through generative AI-based APIs. Nevertheless, it’s identified that using only generative AI doesn’t mean that perfect translation may be achieved. An organization official emphasized, “Perfect translation is set by quality data.”
Flitto has not only learned moreover from the vast amount of knowledge it has collected through the years, but has also recently updated its data by participating in the most important variety of events at home and abroad. It’s also strong in low-resource languages that lack corpus data, akin to Thai and Tagalog.
It is claimed that accuracy has been further improved with a recent update to the generation AI.
Over the course of six months, all processes were meticulously carried out, from ▲problem definition ▲goal setting ▲data collection and model learning ▲service development and verification ▲launch and update. Consequently, it is claimed that they succeeded in significantly complementing the weaknesses of existing translations.
Lee Jeong-soo, CEO of Flitto, said, “Existing machine translation is fast and convenient, but there have been limitations in understanding and translating your complete meaning of a sentence through word or phrase-centered translation.” He added, “To compensate for this, we developed a technology that higher understands the context, akin to cultural nuances and metaphors.” “We added a generative AI-based ‘proofreading (MTPE)’ function,” he said.
Reporter Jang Se-min semim99@aitimes.com
