• Annual Conference on Human and Language Technology
• /
• 2021.10a
• /
• pp.496-497
• /
• 2021

감정 분류는 사람의 사고방식이나 행동양식을 구분하기 위한 중요한 열쇠로, 지난 수십 년간 감정 분석과 관련된 다양한 연구가 진행되었다. 감정 분류의 품질과 정확도를 높이기 위한 방법 중 하나로 단일 레이블링 대신 다중 레이블링된 데이터 세트를 감정 분석에 활용하는 연구가 제안되었고, 본 논문에서는 T5 모델을 한국어와 영어 코퍼스로 학습한 KE-T5 모델을 기반으로 한국어 발화 데이터를 단일 레이블링한 경우와 다중 레이블링한 경우의 감정 분류 성능을 비교한 결과 다중 레이블 데이터 세트가 단일 레이블 데이터 세트보다 23.3% 더 높은 정확도를 보임을 확인했다.

• Journal of the Korea Society of Computer and Information
• /
• v.26 no.7
• /
• pp.1-7
• /
• 2021

In this paper, we propose a feature selection technique for multi-label classification. Many existing feature selection techniques have selected features by calculating the relation between features and labels such as a mutual information scale. However, since the mutual information measure requires a joint probability, it is difficult to calculate the joint probability from an actual premise feature set. Therefore, it has the disadvantage that only a few features can be calculated and only local optimization is possible. Away from this regional optimization problem, we propose a feature selection technique that constructs a low-rank space in the entire given feature space and selects features with sparsity. To this end, we designed a regression-based objective function using Nuclear norm, and proposed an algorithm of gradient descent method to solve the optimization problem of this objective function. Based on the results of multi-label classification experiments on four data and three multi-label classification performance, the proposed methodology showed better performance than the existing feature selection technique. In addition, it was showed by experimental results that the performance change is insensitive even to the parameter value change of the proposed objective function.

• Journal of the Korea Society of Computer and Information
• /
• v.26 no.1
• /
• pp.27-35
• /
• 2021

In this paper, we propose a new stacking ensemble framework for deep learning models which reflects the distribution of label embeddings. Our ensemble framework consists of two phases: training the baseline deep learning classifier, and training the sub-classifiers based on the clustering results of label embeddings. Our framework aims to divide a multi-class classification problem into small sub-problems based on the clustering results. The clustering is conducted on the label embeddings obtained from the weight of the last layer of the baseline classifier. After clustering, sub-classifiers are constructed to classify the sub-classes in each cluster. From the experimental results, we found that the label embeddings well reflect the relationships between classification labels, and our ensemble framework can improve the classification performance on a CIFAR 100 dataset.

• KIPS Transactions on Software and Data Engineering
• /
• v.10 no.12
• /
• pp.555-560
• /
• 2021

Recently, as people's interest in facial skin beauty has increased, research on skin disease recognition for facial skin beauty is being conducted by using deep learning. These studies recognized a variety of skin diseases, including acne. Existing studies can recognize only the single skin diseases, but skin diseases that occur on the face can enact in a more diverse and complex manner. Therefore, in this paper, complex skin diseases such as acne, blackheads, freckles, age spots, normal skin, and whiteheads are identified using the Inception-ResNet V2 deep learning mode with multi-label classification. The accuracy was 98.8%, hamming loss was 0.003, and precision, recall, F1-Score achieved 96.6% or more for each single class.

• Journal of Internet Computing and Services
• /
• v.18 no.1
• /
• pp.77-88
• /
• 2017

Recently, with the advent of knowledge based society where information and knowledge make values, patents which are the representative form of intellectual property have become important, and the number of the patents follows growing trends. Thus, it needs to classify the patents depending on the technological topic of the invention appropriately in order to use a vast amount of the patent information effectively. IPC (International Patent Classification) is widely used for this situation. Researches about IPC automatic classification have been studied using data mining and machine learning algorithms to improve current IPC classification task which categorizes patent documents by hand. However, most of the previous researches have focused on applying various existing machine learning methods to the patent documents rather than considering on the characteristics of the data or the structure of patent documents. In this paper, therefore, we propose to use two structural fields, technical field and background, considered as having impacts on the patent classification, where the two field are selected by applying of the characteristics of patent documents and the role of the structural fields. We also construct multi-label classification model to reflect what a patent document could have multiple IPCs. Furthermore, we propose a method to classify patent documents at the IPC subclass level comprised of 630 categories so that we investigate the possibility of applying the IPC multi-label classification model into the real field. The effect of structural fields of patent documents are examined using 564,793 registered patents in Korea, and 87.2% precision is obtained in the case of using title, abstract, claims, technical field and background. From this sequence, we verify that the technical field and background have an important role in improving the precision of IPC multi-label classification in IPC subclass level.

• The Journal of the Acoustical Society of Korea
• /
• v.42 no.6
• /
• pp.536-543
• /
• 2023

Speaker diarization, which labels for "who spoken when?" in speech with multiple speakers, has been studied on a deep neural network-based end-to-end method for labeling on speech overlap and optimization of speaker diarization models. Most deep neural network-based end-to-end speaker diarization systems perform multi-label classification problem that predicts the labels of all speakers spoken in each frame of speech. However, the performance of the multi-label-based model varies greatly depending on what the threshold is set to. In this paper, it is studied a speaker diarization system using single-label classification so that speaker diarization can be performed without thresholds. The proposed model estimate labels from the output of the model by converting speaker labels into a single label. To consider speaker label permutations in the training, the proposed model is used a combination of Permutation Invariant Training (PIT) loss and cross-entropy loss. In addition, how to add the residual connection structures to model is studied for effective learning of speaker diarization models with deep structures. The experiment used the Librispech database to generate and use simulated noise data for two speakers. When compared with the proposed method and baseline model using the Diarization Error Rate (DER) performance the proposed method can be labeling without threshold, and it has improved performance by about 20.7 %.

• The Journal of Society for e-Business Studies
• /
• v.26 no.1
• /
• pp.29-41
• /
• 2021

Recently, there are increasing attempts to utilize deep learning methodology in the fashion industry. Accordingly, research dealing with various fashion-related problems have been proposed, and superior performances have been achieved. However, the studies for fashion style classification have not reflected the characteristics of the fashion style that one outfit can include multiple styles simultaneously. Therefore, we aim to solve the multi-label classification problem by utilizing the dependencies between the styles. A multi-label recognition model based on a graph convolution network is applied to detect and explore fashion styles' dependencies. Furthermore, we accelerate model training and improve the model's performance through transfer learning. The proposed model was verified by a dataset collected from social network services and outperformed baselines.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2021.07a
• /
• pp.735-736
• /
• 2021

인터넷에서 댓글 시스템은 자신의 의사표현을 위한 시스템으로 널리 사용되고 있다. 하지만 이를 악용하여 상대방에 대한 혐오를 드러내기도 한다. 악성댓글에 대한 적절한 대처를 위해 빠르고 정확한 탐지는 필수적이다. 본 연구에서는 악성 댓글 분류 문제를 해결하기 위해서 순서가 있는 분류 레이블의 성질을 활용한 순서형 회귀 (Ordinal regression) 기반의 분류 모델을 제안한다. 일반적인 분류 모형과는 달리 혐오 발언 정도에 따라 다중 레이블을 부여하여 학습을 진행하였다. 실험을 통해 Korean Hate Speech Dataset에 대해 LSTM기반의 모형의 출력층을 다르게 구성하여 순서형 회귀 기반의 모형들의 성능을 비교하였다. 결과적으로 예측 결과에 대한 조율이 가능한 순서형 회귀 모형이 일반적인 순서형 회귀 모형에 비해서 편향된 예측에 대해 추가적인 성능 향상을 보였다.

• Korean Journal of Remote Sensing
• /
• v.34 no.6_2
• /
• pp.1239-1249
• /
• 2018

Optical remote sensing sensors provide visually more familiar images than radar images. However, it is difficult to discriminate sea ice types in optical images using spectral information based machine learning algorithms. This study addresses two topics. First, we propose a semantic segmentation which is a part of the state-of-the-art deep learning algorithms to identify ice types by learning hierarchical and spatial features of sea ice. Second, we propose a new approach by combining of semi-supervised and active learning to obtain accurate and meaningful labels from unlabeled or unseen images to improve the performance of supervised classification for multiple images. Therefore, we successfully added new labels from unlabeled data to automatically update the semantic segmentation model. This should be noted that an operational system to generate ice type products from optical remote sensing data may be possible in the near future.

• Journal of the Korea Convergence Society
• /
• v.11 no.10
• /
• pp.45-51
• /
• 2020

Recently, many games provide data related to the users' game play, and there have been a few studies that predict opponent move by combining machine learning methods. This study predicts opponent move using match data of a real-time strategy game named ClashRoyale and a multi-label classification based on machine learning. In the initial experiment, binary card properties, binary card coordinates, and normalized time information are input, and card type and card coordinates are predicted using random forest and multi-layer perceptron. Subsequently, experiments were conducted sequentially using the next three data preprocessing methods. First, some property information of the input data were transformed. Next, input data were converted to nested form considering the consecutive card input system. Finally, input data were predicted by dividing into the early and the latter according to the normalized time information. As a result, the best preprocessing step was shown about 2.6% improvement in card type and about 1.8% improvement in card coordinates when nested data divided into the early.