Socar Presents Research Achievements at International Deep Learning Conference 'ICLR 2023'

Socar AI Team. From the left, Manager Won Hyejin, Team Leader Park Kyungho, Manager Kim Hyunsu. (Photo by Socar)

Socar will present its deep learning research achievements at the world’s most prestigious conference in the field of deep learning, ‘ICLR 2023’ workshop.

Socar announced on the 2nd that it will attend the PML4DC workshop operated by ‘ICLR 2023’ on the 5th and present a deep learning research paper. Socar will introduce cases that efficiently train data in environments with limited data resources.

The Socar AI team will present a paper on ‘Analysis of Calibration Effects for Public Intent Classification.’ The paper proposes a method to add calibration to the loss value that occurs when deep learning models train data using the 'cross-entropy' loss, which calculates the difference between actual and predicted values.

The second paper proposed a method for ‘Part-of-Speech Replacement and Feature Space Interpolation for Text Data Augmentation.’ The paper introduces a data augmentation technique that simultaneously applies synonym replacement and feature space interpolation in situations where data is insufficient and effective model training is difficult. By utilizing this technique, sentence classification problems can be solved with excellent performance even in data-scarce situations.

Based on these research achievements, Socar is accelerating improvements in platform operation efficiency. Using natural language data obtained through platform operations, Socar is also developing an AI model that best understands the Socar domain. Socar plans to introduce an AI customer center solution based on this AI model as early as this year.

To this end, Socar is independently researching and developing large language models (LLM) to build conversational AI. The conversational AI understands the customer’s utterance intent and generates appropriate responses, enabling quick and accurate handling of inquiries that arise during vehicle use.

The Socar AI team participating in this paper said, “It is a great opportunity to share Socar’s deep learning research achievements at the world’s most prestigious conference,” and added, “Building on this research, we will continue to study ways to solve and streamline various problems that may occur at Socar using data and technology.”

Meanwhile, Socar is achieving operational efficiency by integrating various data obtained through platforms such as the app and vehicles with AI. In particular, in the vehicle accident sector, AI models automatically process vehicle damage and accident status by utilizing photo data uploaded by customers before and after using Socar, DR-GPS sensor data from vehicle control terminals, and video data from black boxes.

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