Socar Wins Excellence Award at EMNLP 2023, the Premier Conference on Natural Language Processing

Socar AI Team (from left: Team Leader Kyungho Park, Manager Hyunsoo Kim, Manager Chunghyun Jo) (Photo by Socar)

Socar has been recognized for its artificial intelligence (AI) technology capabilities at the world’s leading conference in the field of natural language processing.

Socar announced on the 8th that its AI team (Team Leader Kyungho Park, Managers Hyunsoo Kim, Choonghyun Jo, and Haejin Won) received an excellence award for a paper accepted at ‘EMNLP 2023,’ the most prestigious conference in the field of natural language processing.

EMNLP (Empirical Methods in Natural Language Processing) is regarded as the top international academic conference in the field of natural language processing (NLP). EMNLP covers AI research based on language data, including AI translation, machine reading comprehension, and translation. ‘EMNLP 2023,’ held in Singapore from the 6th to the 10th of this month, featured participation from leading AI companies such as Google DeepMind and Microsoft Research.

The Socar AI team participated in the MRL (Multilingual Representation Learning) workshop held on the 7th, sharing research results on multilingual language processing technologies that can be utilized across various languages. The Socar AI team proposed the ‘Adapt and Prune Strategy for Multilingual Speech Foundation Model on Low-resourced Languages,’ introducing a case where only the necessary parameters are extracted from massive models with billions to hundreds of billions of parameters to ensure performance in new languages or domains.

The method proposed by the Socar AI team first uses a lightweight technique called the 'Lottery Ticket Hypothesis' to extract only the parameters related to the target language from the massive model. The extracted parameters contain grammatical and high-level linguistic features related to the target language, enabling the resolution of multiple problems with only a small number of parameters. Furthermore, the proposed method employs LoRA (Low-Rank Adaptation) to additionally train domain-related knowledge. Through the LoRA technique, training is possible with only about 2% of the parameters, demonstrating that similar performance to existing methods can be achieved with fewer computational resources.

Kyungho Park, Team Leader of Socar AI, said, “It was meaningful to share the speech language foundation model researched at Socar at a conference with the world’s highest authority, and I hope this paper will serve as a reference for research on natural language learning models. We will continue our research efforts to optimize Socar’s mobility services and provide users with more reasonable and convenient mobility experiences.”

Meanwhile, the Socar AI team has published about seven research papers this year at renowned international conferences such as EMNLP and ICLR, and their research outcomes are being applied to various AI products. Notably, the results of this research have been applied to Socar’s upcoming 2024 product, AICC (AI-based Contact Center), helping to reduce the computational resources required to build speech language foundation models.

This content was produced with the assistance of AI translation services.