Domestic AI startup Twigfarm published a press release on the 25th and reported that the hybrid translator developed in-house achieved higher results than Google Translate in the 'BLEU evaluation' that evaluates the quality of the translator. This announcement was interesting because, in fact, AI translation is an area where it is not easy for latecomers to catch up with leading companies.
The success or failure of AI translation quality is largely determined by the size of the training data and the level of the algorithm. In particular, securing actual usage data, which is the most important foundation, is a point of great advantage for first-timers who maintain the service for a long time and have many users. Then, all that remains is the differentiation of the algorithm. How did Twigfarm surpass Google Translate?
I Gave the Dictionary to the AI Translator
The weapon they put forward is the 'translation dictionary'. Twigfarm's hybrid translator analyzes documents to extract commonly used words and sentences, and to decide what terms to translate them into. It means using a kind of 'terminology dictionary' to more accurately translate words that have different meanings in different fields even in the same sentence.
In general, general-purpose AI translators tend to choose the meaning of words that are more probabilistic when translating certain sentences. This is because if it is difficult to specify the field, it is necessary to select more frequently used words to reduce the probability of misinterpretation. For this reason, the performance of translators in specialized fields with relatively low usage is still poor compared to everyday translation.
The fields where Twigfarm scored higher than Google Translator this time were: law, finance, machinery, and medicine. The BLEU evaluation used in the comparison is a method of quantifying the quality of translators by evaluating how similar the results of machine translation and human translation are.
An official from Twigfarm explained in a phone call with a reporter, "In addition to using a separate glossary, our own algorithm that analyzes which dictionary out of numerous glossaries best matches the content in the sentence also had a great impact on this result." . Twigfarm has now completed a patent application for hybrid translation.
The Share of Human Translation Could be Dwindling
Meanwhile, the results of this test mean that there has been another variable introduced in the competition for future translation jobs between humans and machines. Amid the recent increasing quality of translators, professional translation is the area where human translators were expected to outperform machines for a while.
However, as is this case, as technologies to achieve a level comparable to that of everyday translation continue to develop in specialized fields, the possibility of an earlier change in the human and machine translation job structure is also increasing. Of course, the demand for human translators will not disappear completely, but the pie is expected to shrink as the related technology develops.
The translator introduced by Twigfarm this time will be applied to customized API translation services for companies, Gcon Studio, a solution for translation companies, and e-mail-based translation service ‘heybunny.’
[Reporter Lee Geon-han]