• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.171-176
• /
• 2023

최근 등장한 대규모 데이터로 사전학습된 자연어 생성 모델들은 대화 능력 및 코드 생성 태스크등에서 인상적인 성능을 보여주고 있어, 본 논문에서는 대형 언어 모델 (LLM)의 한국어 질문을 SQL 쿼리 (Text-to-SQL) 변환하는 성능을 평가하고자 한다. 먼저, 영어 Text-to-SQL 벤치마크 데이터셋을 활용하여 영어 질의문을 한국어 질의문으로 번역하여 한국어 Text-to-SQL 데이터셋으로 만들었다. 대형 생성형 모델 (GPT-3 davinci, GPT-3 turbo) 의 few-shot 세팅에서 성능 평가를 진행하며, fine-tuning 없이도 대형 언어 모델들의 경쟁력있는 한국어 Text-to-SQL 변환 성능을 확인한다. 또한, 에러 분석을 수행하여 한국어 문장을 데이터베이스 쿼리문으로 변환하는 과정에서 발생하는 다양한 문제와 프롬프트 기법을 활용한 가능한 해결책을 제시한다.

• Journal of the Korea Society of Computer and Information
• /
• v.29 no.2
• /
• pp.197-209
• /
• 2024

This study aimed to analyze recent research trends in Artificial Intelligence (AI) education within South Korea with the overarching objective of exploring the future direction of AI education. For this purpose, an analysis of 697 papers related to AI education published in Research Information Sharing Service (RISS) from 2016 to November 2023 were analyzed using word cloud and Latent Dirichlet Allocation (LDA) topic modeling technique. As a result of the analysis, six major topics were identified: generative AI utilization education, AI ethics education, AI convergence education, teacher perceptions and roles in AI utilization, AI literacy development in university education, and AI-based education and research directions. Based on these findings, I proposed several suggestions, (1) including expanding the use of generative AI in various subjects, (2) establishing ethical guidelines for AI use, (3) evaluating the long-term impact of AI education, (4) enhancing teachers' ability to use AI in higher education, (5) diversifying the curriculum of AI education in universities, (6) analyzing the trend of AI research, and developing an educational platform.

• Journal of the Korea Convergence Society
• /
• v.13 no.1
• /
• pp.51-62
• /
• 2022

AI-based speed prediction studies have been conducted quite actively. However, while the importance of explainable AI is emerging, the study of interpreting and reasoning the AI-based speed predictions has not been carried out much. Therefore, in this paper, 'Explainable Deep Graph Neural Network (GNN)' is devised to analyze the speed prediction and assess the nearby road influence for reasoning the critical contributions to a given road situation. The model's output was explained by comparing the differences in output before and after masking the input values of the GNN model. Using TOPIS traffic speed data, we applied our GNN models for the major congested roads in Seoul. We verified our approach through a traffic flow simulation by adjusting the most influential nearby roads' speed and observing the congestion's relief on the road of interest accordingly. This is meaningful in that our approach can be applied to the transportation network and traffic flow can be improved by controlling specific nearby roads based on the inference results.

• Smart Media Journal
• /
• v.12 no.9
• /
• pp.159-170
• /
• 2023

It takes a lot of time and manpower to search for the missing. As part of the solution, a missing person search AI system was implemented using a YOLO-based model. In order to train object detection models, the model was learned by collecting recognition images (road fixation) of drone mobile objects from AI-Hub. Additional mountainous terrain datasets were also collected to evaluate performance in training datasets and other environments. In order to optimize the missing person search AI system, performance evaluation based on model size and hyperparameters and additional performance evaluation for concerns about overfitting were conducted. As a result of performance evaluation, it was confirmed that the YOLOv5-L model showed excellent performance, and the performance of the model was further improved by applying data augmentation techniques. Since then, the web service has been applied with the YOLOv5-L model that applies data augmentation techniques to increase the efficiency of searching for missing people.

• Informatization Policy
• /
• v.30 no.3
• /
• pp.69-91
• /
• 2023

The impact of technological advancements on employment is a matter of ongoing debate, with discussions on the effects of AI technology development on employment being particularly scarce. This study employs the natural language processing technique (SBERT) and patents to calculate an occupation-based AI exposure score and to analyze employment trends by group. It proposes a method for calculating the AI exposure score based on the similarity between Korean patent information and US job descriptions and linking SOC(U.S.) and KSCO(Korea). The analysis of domestic AI patent applications and regional employment data in the KOSIS Database since 2013 reveals a K-shaped polarization pattern in Korean employment trends among groups with above and below average levels of AI exposure.

• Water for future
• /
• v.29 no.5
• /
• pp.123-125
• /
• 1996

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.33 no.4
• /
• pp.621-631
• /
• 2023

Adversarial training improves the robustness of deep neural networks for adversarial examples. However, the previous adversarial training method focuses only on the adversarial loss function, ignoring that even a small perturbation of the input layer causes a significant change in the hidden layer features. Consequently, the accuracy of a defended model is reduced for various untrained situations such as clean samples or other attack techniques. Therefore, an architectural perspective is necessary to improve feature representation power to solve this problem. In this paper, we apply an attention module that generates an attention map of an input image to a general model and performs PGD adversarial training upon the augmented model. In our experiments on the CIFAR-10 dataset, the attention augmented model showed higher accuracy than the general model regardless of the network structure. In particular, the robust accuracy of our approach was consistently higher for various attacks such as PGD, FGSM, and BIM and more powerful adversaries. By visualizing the attention map, we further confirmed that the attention module extracts features of the correct class even for adversarial examples.

• Journal of Korea Society of Industrial Information Systems
• /
• v.23 no.2
• /
• pp.63-73
• /
• 2018

Academia and research institutes mainly use qualitative methods that rely on expert judgments to understand and predict research trends and science and technology trends. Since such a technique has the disadvantage of requiring much time and money, in this study, science and technology trends were predicted using keyword network analysis. To that end, 13,618 AI (Artificial Intelligence) patent abstracts were analyzed using keyword network analysis in three separate lots based on the period of the submission of each abstract: analysis period 1 (January 1, 2002 - December 31, 2006), analysis period 2 (January 1, 2007 - December 31, 2011), and analysis period 3 (January 1, 2012 - December 31, 2016). According to the results of frequency analyses, keywords related to methods in the field of AI application appeared more frequently as time passed from analysis period 1 to analysis period 3. In keyword network analyses, the connectivity between keywords related to methods in the field of AI application and other keywords increased over time. In addition, when the connected keywords that showed increasing or decreasing trends during the entire analysis period were analyzed, it could be seen that the connectivity to methods and management in the field of AI application was strengthened while the connectivity to the field of basic science and technology was weakened. According to analysis of keyword connection centrality, the centrality value of the field of AI application increased over time. According to analysis of keyword mediation centrality during analysis period 3, keywords related to methodologies in the field of AI application showed the highest mediation value. Therefore, it is expected that methods in the field of AI application will play the role of powerful intermediaries in AI hereafter. The technique presented in this paper can be employed in the excavation of tasks related to regional innovation or in fields such as social issue visualization.

• Annual Conference on Human and Language Technology
• /
• 2022.10a
• /
• pp.82-87
• /
• 2022

대화 모델은 대표적으로 검색 모델 또는 생성 모델을 기반으로 구현된다. 최근에는 두 모델의 장점은 융합하고 단점은 보완하기 위해 검색 기법과 생성 기법을 결합하는 연구가 활발히 이루어지고 있다. 그러나 생성 모델이 검색된 응답을 전혀 반영하지 않고 응답을 생성하여 검색 모델을 간과하는 문제 또는 검색된 응답을 그대로 복사해 생성하여 검색 모델에 과의존하는 문제가 발생한다. 본 논문에서는 이러한 문제들을 완화하며 검색 모델과 생성 모델을 모두 조화롭게 활용할 수 있는 대화 모델을 제안한다. 생성 모델이 검색 모델을 간과하는 문제를 완화하기 위해 학습 시 골드 응답을 검색된 응답과 함께 사용한다. 또한, 검색 모델에 과의존하는 문제를 완화하기 위해 검색된 응답들의 내용어 일부를 마스킹하고 순서를 무작위로 섞어 노이징한다. 검색된 응답은 대화 컨텍스트와의 관련성이 높은 것만을 선별하여 생성에 활용한다. 정량 평가 및 정성 평가를 통해 제안한 방법의 성능 향상 효과를 확인하였다.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.05a
• /
• pp.603-605
• /
• 2023

최근 SNS 의 발달로 인해 자신의 감정을 빠르고 효과적으로 전달할 수 있는 이모지의 중요성이 커지고 있다. 하지만 이모지를 수동으로 생성하기 위해서 시간과 비용이 많이 들고 자신의 감정에 맞는 이모지를 찾아야 하며 해당 이모지가 없을 수 있다. 기존 DCGAN 을 활용한 이모지 자동 생성연구에서는 부족한 데이터셋으로 인해 G(Generator)와 D(Discriminator)가 동등하게 학습하지 못해서 두 모델 간 성능 차이가 발생한다. D 가 G 보다 최적해에 빠르게 수렴하여 G 가 학습이 되지 않아 낮은 품질의 이모지를 생성하는 불안정 문제가 발생한다. 이 문제를 해결하기 위해 본 논문에서는 Latent vector 분포를 데이터셋에 맞게 조정하여 적은 데이터로 G 에서 안정적으로 학습할 수 있게 하는 G 구조와 다양한 이모지 생성을 위한 Latent vector 평균 조정 기법을 제안한다. 비교 실험 결과 불안정 문제를 개선하였고 FID 와 IS 수치를 통해 성능 개선 효과를 검증했다.