Kakao Enterprise ranked first in the global English machine reading competition, beating leading AI powerhouses such as the United States and China. After achieving first place in the Korean machine reading evaluation ‘KorQuAD’ in 2019 and 2022, it also proved its global artificial intelligence (AI) technology in the English machine reading competition ELI5 (Explain Like I’m 5).
ELI5 is a global machine reading competition hosted by Meta AI’s Knowledge Integrated Language Tasks (KILT) integrated benchmark. It goes beyond extracting answers from learned documents by referring to multiple documents. Evaluate the performance of generating answers in sentence format. Because it is a competition targeting models that can directly create text that does not exist in a document, it is more difficult than a simple extraction competition.
Generative models are a field in which research has been ongoing relatively recently compared to other models in the field of machine reading. Kakao Enterprise was the first domestic company to participate in this competition and took first place, further solidifying its status as Korea’s leading AI company.
This model is an improved version of the ‘LittleBird’ model, which ranked first in Coquad, to fit a generative model. It is noteworthy that while most companies submit large models to achieve high performance, Kakao Enterprise took first place with a lightweight model optimized for speed. The model submitted by Kakao Enterprise is easier to apply to actual services compared to large models, and is advantageous in terms of cost and efficiency in that it achieves global top performance even with a small model.
Based on this model, Kakao Enterprise plans to further enhance its services such as AI contact center ‘Kakao i Connect Center’ and AI chatbot ‘Kakao i Connect Talk’. Previously, it was necessary to classify answers to questions one by one, but this model makes it possible to automate learning from documents and directly generating answers. This is suitable for companies whose response manuals change frequently and require quick service application.
Choi Dong-jin, Chief Artificial Intelligence Officer (CAIO) and Vice President of Kakao Enterprise, said, “This first place in ELI5 has significant implications for the AI industry in that it demonstrated global first-place performance with a lightweight model optimized for speed. “We will gradually expand the model and create a model that combines both performance and efficiency,” he said. “Kakao Enterprise will enable AI contact centers and AI chatbots to accurately deliver the information needed to users as well as generate answers in more natural sentences.” “We plan to continue research,” he said.
Meanwhile, Kakao Enterprise has ranked first twice in Coquad, an AI Korean reading ability evaluation, in March 2019 and 2022. In 2019, Coquad 1.0 recorded a score 0.65 points higher than the 91.20 points stated to be the level of human sentence discrimination and reading ability, making it the first AI to score higher than humans in the field of Korean machine reading. Additionally, in 2021, it took first place in the ‘Passage Ranking’ category of the global AI natural language understanding competition ‘MS MARCO’. In addition, the company is continuing its path as Korea’s representative AI company by taking first place in consecutive years at FRVT, a global AI facial recognition competition, in May and June and presenting four papers at InterSpeech, a global speech processing conference, in September.
Source: Pangyo Techno Valley Official Newsroom
→ Go to ‘Asian Innovation Hub Pangyo Techno Valley 2022’ news