• KSII Transactions on Internet and Information Systems (TIIS)
• /
• 제12권11호
• /
• pp.5287-5303
• /
• 2018

Handling vast amount of data found in large transactional datasets is an obvious challenge for the conventional data mining algorithms. Addressing this challenge, our paper proposes a parallel approach for proper decomposition of mining problem into sub-problems in order to find frequent patterns from these datasets. The proposed, Pattern Mining for Large Distributed Dataset (PMLDD) approach, ensures minimum dependencies as well as minimum communications among sub-problems. It establishes a linear aggregation of the intermediate results so that it can be adapted to large-scale programming models like MapReduce. In this context, an algorithmic structure for MapReduce programming model is presented. PMLDD guarantees an efficient load balancing among the sub-problems by a specific selection criterion. Further, it optimizes the number of required iterations over the dataset for mining frequent patterns as compared to the existing approaches. Finally, we believe that our approach is scalable enough to handle larger datasets in terms of performance evaluation, and the result analysis justifies all these mentioned concerns.

• 대한임베디드공학회논문지
• /
• 제17권5호
• /
• pp.273-280
• /
• 2022

Along with the advancement of deep learning technology, securing high-quality dataset for verification of developed technology is emerging as an important issue, and developing robust deep learning models to the domestic road environment is focused by many research groups. Especially, unlike expressways and automobile-only roads, in the complex city driving environment, various dynamic objects such as motorbikes, electric kickboards, large buses/truck, freight cars, pedestrians, and traffic lights are mixed in city road. In this paper, we built our dataset through multi camera-based processing (collection, refinement, and annotation) including the various objects in the city road and estimated quality and validity of our dataset by using YOLO-based model in object detection. Then, quantitative evaluation of our dataset is performed by comparing with the public dataset and qualitative evaluation of it is performed by comparing with experiment results using open platform. We generated our 2D dataset based on annotation rules of KITTI/COCO dataset, and compared the performance with the public dataset using the evaluation rules of KITTI/COCO dataset. As a result of comparison with public dataset, our dataset shows about 3 to 53% higher performance and thus the effectiveness of our dataset was validated.

• 한국환경보건학회지
• /
• 제45권2호
• /
• pp.192-202
• /
• 2019

Objectives: According to the central limit theorem, the samples in population might be considered to follow normal distribution if a large number of samples are available. Once we assume that toxicity dataset follow normal distribution, we can treat and process data statistically to calculate genus or species mean value with standard deviation. However, little is known and only limited studies are conducted to investigate whether toxicity dataset follows normal distribution or not. Therefore, the purpose of study is to evaluate the generally accepted normality hypothesis of aquatic toxicity dataset Methods: We selected the 8 chemicals, which consist of 4 organic and 4 inorganic chemical compounds considering data availability for the development of species sensitivity distribution. Toxicity data were collected at the US EPA ECOTOX Knowledgebase by simple search with target chemicals. Toxicity data were re-arranged to a proper format based on the endpoint and test duration, where we conducted normality test according to the Shapiro-Wilk test. Also we investigated the degree of normality by simple log transformation of toxicity data Results: Despite of the central limit theorem, only one large dataset (n>25) follow normal distribution out of 25 large dataset. By log transforming, more 7 large dataset show normality. As a result of normality test on small dataset (n<25), log transformation of toxicity value generally increases normality. Both organic and inorganic chemicals show normality growth for 26 species and 30 species, respectively. Those 56 species shows normality growth by log transformation in the taxonomic groups such as amphibian (1), crustacean (21), fish (22), insect (5), rotifer (2), and worm (5). In contrast, mollusca shows normality decrease at 1 species out of 23 that originally show normality. Conclusions: The normality of large toxicity dataset was not always satisfactory to the central limit theorem. Normality of those data could be improved through log transformation. Therefore, care should be taken when using toxicity data to induce, for example, mean value for risk assessment.

• 정보처리학회 논문지
• /
• 제13권4호
• /
• pp.181-188
• /
• 2024

기존의 비속어 탐지 방법들은 의도적으로 변형된 비속어를 식별하는 데 한계가 있다. 이 논문에서는 자연어 처리의 한 분야인 개체명 인식에 기반한 새로운 방법을 소개한다. 우리는 시퀀스 레이블링을 이용한 비속어 탐지 기법을 개발하고, 이를 위해 한국어 악성 댓글 중 일부 비속어를 레이블링하여 직접 데이터셋을 구축하여 실험을 수행하였다. 또한 모델의 성능을 향상시키기 위하여 거대 언어 모델중 하나인 ChatGPT를 활용해 한국어 혐오발언 데이터셋의 일부를 레이블링을 하는 방식으로 데이터셋을 증강하여 학습을 진행하였고, 이 과정에서 거대 언어 모델이 생성한 데이터셋을 인간이 필터링 하는 것만으로도 성능을 향상시킬 수 있음을 확인하였다. 이를 통해 데이터셋 증강 과정에는 여전히 인간의 관리감독이 필요함을 제시하였다.

• 정보처리학회논문지:소프트웨어 및 데이터공학
• /
• 제12권11호
• /
• pp.481-492
• /
• 2023

본 연구는 생성형 대규모 언어 모델을 활용하여 텍스트에서 정보를 추출하기 위한 한글 데이터셋 구축 방법을 탐구한다. 현대 사회에서는 혼합된 정보가 빠르게 유포되며, 이를 효과적으로 분류하고 추출하는 것은 의사결정 과정에 중요하다. 그러나 이에 대한 학습용 한국어 데이터셋은 아직 부족하다. 이를 극복하기 위해, 본 연구는 생성형 대규모 언어 모델을 사용하여 텍스트 기반 제로샷 학습(zero-shot learning)을 이용한 정보 추출을 시도하며, 이를 통해 목적에 맞는 한국어 데이터셋을 구축한다. 본 연구에서는 시스템-지침-소스입력-출력형식의 프롬프트 엔지니어링을 통해 언어 모델이 원하는 결과를 출력하도록 지시하며, 입력 문장을 통해 언어 모델의 In-Context Learning 특성을 활용하여 데이터셋을 구축한다. 생성된 데이터셋을 기존 데이터셋과 비교하여 본 연구 방법론을 검증하며, 관계 정보 추출 작업의 경우 KLUE-RoBERTa-large 모델 대비 25.47% 더 높은 성능을 달성했다. 이 연구 결과는 한국어 텍스트에서 지식 요소를 추출하는 가능성을 제시함으로써 인공지능 연구에 도움을 줄 것으로 기대된다. 더욱이, 이 방법론은 다양한 분야나 목적에 맞게 활용될 수 있어, 다양한 한국어 데이터셋 구축에 잠재력을 가진다고 볼 수 있다.

• 방송공학회논문지
• /
• 제28권3호
• /
• pp.302-313
• /
• 2023

본 연구는 과학기술정보통신부가 2017년부터 1조원 이상을 투자한 'AI Hub 댐' 사업에서 구축된 인공지능 모델 학습데이터의 품질관리를 자동화할 수 있는 프레임워크의 개발을 목표로 한다. 자율주행 개발에 사용되는 AI 모델 학습에는 다량의 고품질의 데이터가 필요하며, 가공된 데이터를 검수자가 데이터 자체의 이상을 검수하고 유효함을 증명하는 데는 여전히 어려움이 있으며 오류가 있는 데이터로 학습된 모델은 실제 상황에서 큰 문제를 야기할 수 있다. 본 논문에서는 이상 데이터를 제거하는 신뢰할 수 있는 데이터셋 정제 프레임워크를 통해 모델의 인식 성능을 향상시키는 전략을 소개한다. 제안하는 방법은 인공지능 학습용 데이터 품질관리 가이드라인의 지표를 기반으로 설계되었다. 한국정보화진흥원의 AI Hub을 통해 공개된 자율주행 데이터셋에 대한 실험을 통해 프레임워크의 유효성을 증명하였고, 이상 데이터가 제거된 신뢰할 수 있는 데이터셋으로 재구축될 수 있음을 확인하였다.

• 말소리와 음성과학
• /
• 제15권3호
• /
• pp.83-88
• /
• 2023

This study investigates the fine-tuning of large-scale Automatic Speech Recognition (ASR) models, specifically OpenAI's Whisper model, for domain-specific applications using the KsponSpeech dataset. The primary research questions address the effectiveness of targeted lexical item emphasis during fine-tuning, its impact on domain-specific performance, and whether the fine-tuned model can maintain generalization capabilities across different languages and environments. Experiments were conducted using two fine-tuning datasets: Set A, a small subset emphasizing specific lexical items, and Set B, consisting of the entire KsponSpeech dataset. Results showed that fine-tuning with targeted lexical items increased recognition accuracy and improved domain-specific performance, with generalization capabilities maintained when fine-tuned with a smaller dataset. For noisier environments, a trade-off between specificity and generalization capabilities was observed. This study highlights the potential of fine-tuning using minimal domain-specific data to achieve satisfactory results, emphasizing the importance of balancing specialization and generalization for ASR models. Future research could explore different fine-tuning strategies and novel technologies such as prompting to further enhance large-scale ASR models' domain-specific performance.

• 대한임베디드공학회논문지
• /
• 제18권5호
• /
• pp.203-208
• /
• 2023

Recently, with the growing interest in the field of autonomous driving, many researchers have been focusing on developing autonomous driving software platforms. In particular, we have concentrated on developing 3D object detection models that can improve real-time performance. In this paper, we introduce a self-constructed 3D LiDAR dataset specific to domestic environments and propose a VariFocal-based CenterPoint for the 3D object detection model, with improved performance over the previous models. Furthermore, we present experimental results comparing the performance of the 3D object detection modules using our self-built and public dataset. As the results show, our model, which was trained on a large amount of self-constructed dataset, successfully solves the issue of failing to detect large vehicles and small objects such as motorcycles and pedestrians, which the previous models had difficulty detecting. Consequently, the proposed model shows a performance improvement of about 1.0 mAP over the previous model.

• Journal of Information Processing Systems
• /
• 제19권5호
• /
• pp.602-613
• /
• 2023

High-quality image datasets are in high demand for various applications. With many online sources providing manually collected datasets, a persisting challenge is to fully automate the dataset collection process. In this study, we surveyed an automatic image dataset generation field through analyzing a collection of existing studies. Moreover, we examined fields that are closely related to automated dataset generation, such as query expansion, web scraping, and dataset quality. We assess how both noise and regional search engine differences can be addressed using an automated search query expansion focused on hypernyms, allowing for user-specific manual query expansion. Combining these aspects provides an outline of how a modern web scraping application can produce large-scale image datasets.

• International Journal of Internet, Broadcasting and Communication
• /
• 제16권3호
• /
• pp.27-32
• /
• 2024

Reliable and fine-grained musical metadata are required for efficient search of rapidly increasing music files. In particular, since the primary motive for listening to music is its emotional effect, diversion, and the memories it awakens, emotion classification along with genre classification of music is crucial. In this paper, as an initial approach towards a "ground-truth" dataset for music emotion and genre classification, we elaborately generated a music corpus through labeling of a large number of ordinary people. In order to verify the suitability of the dataset through the classification results, we extracted features according to MPEG-7 audio standard and applied different machine learning models based on statistics and deep neural network to automatically classify the dataset. By using standard hyperparameter setting, we reached an accuracy of 93% for genre classification and 80% for emotion classification, and believe that our dataset can be used as a meaningful comparative dataset in this research field.