loaction
SKT AI 2nd Fellowship, SNU Team of Graduate School of Industrial Engineering Selected as the Most Outstanding Team for Research on the Improvement of AI Capabilities
-
작성자
관리자
-
등록일
2021.01.15
-
조회수
1,111
SKT AI 2nd Fellowship, SNU Team of Graduate School of Industrial Engineering Selected as the Most Outstanding Team for Research on the Improvement of AI Capabilities
-Drew attention on the research results that led to the dramatic improvement in performance of summarizing abilities in Korean through deep learning technology
- From being utilized in research papers to being used in customer consultation logs, there is an infinite possibility of applying the research results
(From left) SNU College of Engineering Industrial Engineering Data Mining Center M.S. students Seoyoung Pak, Jimin Sun and M.S./Ph.D. student Jaewon Lee
- From being utilized in research papers to being used in customer consultation logs, there is an infinite possibility of applying the research results
(From left) SNU College of Engineering Industrial Engineering Data Mining Center M.S. students Seoyoung Pak, Jimin Sun and M.S./Ph.D. student Jaewon Lee
SNU College of Engineering (Dean Kookheon Char) announced on December 11 that M.S./Ph.D. student Jaewon Lee and M.S. students Seoyoung Pak and Jimin Sun of the Data Mining Center (Advisor: Professor Sungzoon Cho) of the Department of Industrial Engineering were selected as the most outstanding team of SK Telecom's 2nd AI Fellowship.
The SKT AI Fellowship is a program by SK Telecom that provides practical experience and research opportunities in major ICT fields such as AI and 5G tp undergraduate and graduate students across the country. It was started in 2019 and this is its 2nd year. A total of six teams were selected to be part of the 2nd SKT AI Fellowship Program through two screening processes in May, and for the next 6 months, they performed tasks on topics regarding major ICT sectors such as AI, Big Data, Mobility and Security.
The theme for the three researchers that were selected as the most outstanding team was "Korean Summarization Model & Dataset," and they conducted their research with SK Telecom's mentor (Research Engineer Heewon Jeon, SK Telecom's AI Language Tech. Labs).
The researchers used a model that utilized KoBERT, a Korean deep learning technology, to solve the problem of "summarization" and established a Korean summarizing dataset that is required for model learning. The groundbreaking BERT model announced by Google in 2018 changed the landscape of natural language processing. In 2019, SK Telecom developed KoBERT, a BERT model specialized in Korean, and released it as an open source, which contributed to the development of various Korean natural language processing tasks.
The "summarizing" technique can be applied to lengthy documents such as legal documents or research papers as well as to relatively short documents like product reviews and customer consultation logs, showing that there are infinite ways to utilize it. Korean summarization is one of the natural language processing fields that has been relatively less studied than English summarization, and shows great potential for development.
The final outcome will be released as an open source after being reviewed by SK Telecom. "Through this study, we wish to contribute to the research and development of Korean language summarization and take a step further and contribute in the processing of Korean natural language," said the researchers.
The final findings of the three researchers are as follows.
1) Implementation of 3 Single-document Abstractive Models to be fitted into Korean, which shows to be capable of generating a new, human-like summary of a document. A combined extraction of news article summaries from different media outlets covering the same topic was utilized to build the datasets needed for learning in the three models.
2) Implementation of a Multi-document Extraction Model (TextRank+MMR) that imports the most important phrases from multiple documents. The method of topic-based grouping of news articles from various media outlets was utilized to build a multi-document summary dataset for learning in this model.
Being selected as the most outstanding team, the research team has also earned the opportunity to present its research results to all members of SK Telecom. "We sincerely thank our mentor Heewon Jeon, SK Telecom's Competency Culture Group Manager, our advisor Sungzoon Cho as well as all the senior and junior researchers at the Data Mining Center who supported us with their generous encouragements and support throughout the Fellowship," said the three researchers.