• Annual Conference on Human and Language Technology
• /
• 2020.10a
• /
• pp.237-240
• /
• 2020

사전학습 언어모델은 다양한 자연어처리 작업에서 높은 성능을 보였다. 하지만, 사전학습 언어모델은 문장 내 문맥 정보만을 학습하기 때문에 단어간 의미관계 정보를 추론하는데는 한계가 있다. 최근에는, 사전학습 언어모델이 어느수준으로 단어간 의미관계를 이해하고 있는지 다양한 Probing Test를 진행하고 있다. 이러한 Test는 언어모델의 강점과 약점을 분석하는데 효율적이며, 한층 더 인간의 언어를 정확하게 이해하기 위한 모델을 구축하는데 새로운 방향을 제시한다. 본 논문에서는 대표적인 사전 학습기반 언어모델인 BERT(Bidirectional Encoder Representations from Transformers)의 단어간 의미관계 이해도를 평가하는 3가지 작업을 진행한다. 첫 번째로 단어 간의 상위어, 하위어 관계를 나타내는 IsA 관계를 분석한다. 두번째는 '자동차'와 '변속'과 같은 관계를 나타내는 PartOf 관계를 분석한다. 마지막으로 '새'와 '날개'와 같은 관계를 나타내는 HasA 관계를 분석한다. 결과적으로, BERT_base 모델에 대해서는 추론 결과 대부분에서 낮은 성능을 보이지만, BERT_large 모델에서는 BERT_base보다 높은 성능을 보였다.

• Annual Conference on Human and Language Technology
• /
• 2020.10a
• /
• pp.145-150
• /
• 2020

최근 음성인식 분야에서 신경망 기반의 종단간 모델이 제안되고 있다. 해당 모델들은 음성을 직접 입력받아 전사된 문장을 생성한다. 음성을 직접 입력받는 모델의 특성상 데이터의 품질이 모델의 성능에 많은 영향을 준다. 본 논문에서는 이러한 종단간 모델의 문제점을 해결하고자 음성인식 결과를 후처리하기 위한 멀티모달 기반 모델을 제안한다. 제안 모델은 음성과 전사된 문장을 입력 받는다. 입력된 각각의 데이터는 Encoder를 통해 자질을 추출하고 주의집중 메커니즘을 통해 Decoder로 추출된 정보를 전달한다. Decoder에서는 전달받은 주의집중 메커니즘의 결과를 바탕으로 후처리된 토큰을 생성한다. 본 논문에서는 후처리 모델의 성능을 평가하기 위해 word error rate를 사용했으며, 실험결과 Google cloud speech to text모델에 비해 word error rate가 8% 감소한 것을 확인했다.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2022.11a
• /
• pp.55-56
• /
• 2022

Hypertension is a severe health problem and increases the risk of other health issues, such as heart disease, heart attack, and stroke. In this research, we propose a machine learning-based prediction method for the risk of chronic hypertension. The proposed method consists of four main modules. In the first module, the linear interpolation method fills missing values of the integration of gas and meteorological datasets. In the second module, the OrdinalEncoder-based normalization is followed by the Decision tree algorithm to select important features. The prediction analysis module builds three models based on k-Nearest Neighbors, Decision Tree, and Random Forest to predict hypertension levels. Finally, the features used in the prediction model are explained by the DeepSHAP approach. The proposed method is evaluated by integrating the Korean meteorological agency dataset, natural gas leakage dataset, and Korean National Health and Nutrition Examination Survey dataset. The experimental results showed important global features for the hypertension of the entire population and local components for particular patients. Based on the local explanation results for a randomly selected 65-year-old male, the effect of hypertension increased from 0.694 to 1.249 when age increased by 0.37 and gas loss increased by 0.17. Therefore, it is concluded that gas loss is the cause of high blood pressure.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2022.07a
• /
• pp.123-125
• /
• 2022

Git의 커밋 메시지는 프로젝트가 진행되면서 발생하는 각종 이슈 및 코드의 변경이력을 저장하고 관리하고 있기 때문에 소프트웨어 유지관리와 프로젝트의 생명주기와 밀접한 연관성을 갖고 있다. 이러한 Git의 커밋 메시지에 대한 정확한 분석 결과는 소프트웨어 개발 및 유지관리 활동 시, 시간과 비용의 효율적인 관리에 많은 영향을 끼치고 있다. 이에 대한 기존 연구로 Git에서 발생하는 커밋 메시지를 소프트웨어 유지관리의 세 가지 형태로 분류하고 매핑하여 정확한 분석을 시도하려는 연구가 진행되었으나, 최대 87%의 정확도를 제시한 연구 결과가 있었다. 이러한 연구들은 정확도가 낮아 실제 프로젝트의 개발 및 유지관리에 적용하기에는 위험성과 어려움이 있는 현실이다. 본 논문에서는 커밋 메시지 분류에 대한 선행 연구 조사를 통해 각 연구들의 프로세스와 특징을 추출하였고, 이를 이용한 분류 정확도를 높일 수 있는 커밋 복합 분류 모델에 대해 제안한다.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2022.06a
• /
• pp.1191-1194
• /
• 2022

인터넷의 발전으로 수많은 이미지와 비디오를 손쉽게 이용할 수 있게 되었다. 이미지와 비디오 데이터의 양이 기하급수적으로 증가함에 따라, JPEG, HEVC, VVC 등 이미지와 비디오를 효율적으로 저장하기 위한 부호화 기술들이 등장했다. 최근에는 인공신경망을 활용한 학습 기반 모델이 발전함에 따라, 이를 활용한 이미지 및 비디오 압축 기술에 관한 연구가 빠르게 진행되고 있다. NNIC (Neural Network based Image Coding)는 이러한 학습 가능한 인공신경망 기반 이미지 부호화 기술을 의미한다. 본 논문에서는 NNIC 모델과 인공신경망 기반의 초해상화(Super Resolution) 모델을 합동훈련하여 기존 NNIC 모델보다 더 높은 성능을 보일 수 있는 방법을 제시한다. 먼저 NNIC 인코더(Encoder)에 이미지를 입력하기 전 다운 스케일링(Down Scaling)으로 쌍삼차보간법을 사용하여 이미지의 화소를 줄인 후 부호화(Encoding)한다. NNIC 디코더(Decoder)를 통해 부호화된 이미지를 복호화(Decoding)하고 업 스케일링으로 초해상화를 통해 복호화된 이미지를 원본 이미지로 복원한다. 이때 NNIC 모델과 초해상화 모델을 합동훈련한다. 결과적으로 낮은 비트량에서 더 높은 성능을 볼 수 있는 가능성을 보았다. 또한 합동훈련을 함으로써 전체 성능의 향상을 보아 학습 시간을 늘리고, 압축 잡음을 위한 초해상화 모델을 사용한다면 기존의 NNIC 보다 나은 성능을 보일 수 있는 가능성을 시사한다.

• ETRI Journal
• /
• v.46 no.1
• /
• pp.82-95
• /
• 2024

As automated essay scoring (AES) has progressed from handcrafted techniques to deep learning, holistic scoring capabilities have merged. However, specific trait assessment remains a challenge because of the limited depth of earlier methods in modeling dual assessments for holistic and multi-trait tasks. To overcome this challenge, we explore providing comprehensive feedback while modeling the interconnections between holistic and trait representations. We introduce the DualBERT-Trans-CNN model, which combines transformer-based representations with a novel dual-scale bidirectional encoder representations from transformers (BERT) encoding approach at the document-level. By explicitly leveraging multi-trait representations in a multi-task learning (MTL) framework, our DualBERT-Trans-CNN emphasizes the interrelation between holistic and trait-based score predictions, aiming for improved accuracy. For validation, we conducted extensive tests on the ASAP++ and TOEFL11 datasets. Against models of the same MTL setting, ours showed a 2.0% increase in its holistic score. Additionally, compared with single-task learning (STL) models, ours demonstrated a 3.6% enhancement in average multi-trait performance on the ASAP++ dataset.

• KIPS Transactions on Software and Data Engineering
• /
• v.12 no.8
• /
• pp.341-354
• /
• 2023

The title of a document is the brief summarization of the document. Readers can easily understand a document if we provide them with its title in their preferred styles and the languages. In this research, we propose a cross-lingual and style-based title generation model using multiple adapters. To train the model, we need a parallel corpus in several languages with different styles. It is quite difficult to construct this kind of parallel corpus; however, a monolingual title generation corpus of the same style can be built easily. Therefore, we apply a zero-shot strategy to generate a title in a different language and with a different style for an input document. A baseline model is Transformer consisting of an encoder and a decoder, pre-trained by several languages. The model is then equipped with multiple adapters for translation, languages, and styles. After the model learns a translation task from parallel corpus, it learns a title generation task from monolingual title generation corpus. When training the model with a task, we only activate an adapter that corresponds to the task. When generating a cross-lingual and style-based title, we only activate adapters that correspond to a target language and a target style. An experimental result shows that our proposed model is only as good as a pipeline model that first translates into a target language and then generates a title. There have been significant changes in natural language generation due to the emergence of large-scale language models. However, research to improve the performance of natural language generation using limited resources and limited data needs to continue. In this regard, this study seeks to explore the significance of such research.

• KIPS Transactions on Software and Data Engineering
• /
• v.12 no.12
• /
• pp.505-518
• /
• 2023

3D point cloud semantic segmentation is a computer vision task that involves dividing the point cloud into different objects and regions by predicting the class label of each point. Existing 3D semantic segmentation models have some limitations in performing sufficient fusion of multi-modal features while ensuring both characteristics of 2D visual features extracted from RGB images and 3D geometric features extracted from point cloud. Therefore, in this paper, we propose MMCA-Net, a novel 3D semantic segmentation model using 2D-3D multi-modal features. The proposed model effectively fuses two heterogeneous 2D visual features and 3D geometric features by using an intermediate fusion strategy and a multi-modal cross attention-based fusion operation. Also, the proposed model extracts context-rich 3D geometric features from input point cloud consisting of irregularly distributed points by adopting PTv2 as 3D geometric encoder. In this paper, we conducted both quantitative and qualitative experiments with the benchmark dataset, ScanNetv2 in order to analyze the performance of the proposed model. In terms of the metric mIoU, the proposed model showed a 9.2% performance improvement over the PTv2 model using only 3D geometric features, and a 12.12% performance improvement over the MVPNet model using 2D-3D multi-modal features. As a result, we proved the effectiveness and usefulness of the proposed model.

• KIPS Transactions on Software and Data Engineering
• /
• v.12 no.12
• /
• pp.519-524
• /
• 2023

In hashing-based image retrieval, the hash code of a manipulated image is different from the original image, making it difficult to search for the same image. This paper proposes and evaluates a self-supervised deephashing model that generates perceptual hash codes from feature information such as texture, shape, and color of images. The comparison models are autoencoder-based variational inference models, but the encoder is designed with a fully connected layer, convolutional neural network, and transformer modules. The proposed model is a variational inference model that includes a SimAM module of extracting geometric patterns and positional relationships within images. The SimAM module can learn latent vectors highlighting objects or local regions through an energy function using the activation values of neurons and surrounding neurons. The proposed method is a representation learning model that can generate low-dimensional latent vectors from high-dimensional input images, and the latent vectors are binarized into distinguishable hash code. From the experimental results on public datasets such as CIFAR-10, ImageNet, and NUS-WIDE, the proposed model is superior to the comparative model and analyzed to have equivalent performance to the supervised learning-based deephashing model. The proposed model can be used in application systems that require low-dimensional representation of images, such as image search or copyright image determination.

• Journal of the Institute of Electronics Engineers of Korea SP
• /
• v.44 no.2 s.314
• /
• pp.1-11
• /
• 2007

H.264/AVC uses variable block sizes to achieve significant coding gain. It has 7 different coding modes having different motion compensation block sizes in Inter slice, and 2 different intra prediction modes in Intra slice. This fine-tuned new coding feature has achieved far more significant coding gain compared with previous video coding standards. However, extremely high computational complexity is required when rate-distortion optimization (RDO) algorithm is used. This computational complexity is a major problem in implementing real-time H.264/AVC encoder on computationally constrained devices. Therefore, there is a clear need for complexity reduction algorithm of H.264/AVC such as fast mode decision. In this paper, we propose a fast mode decision with early $P8\times8$ mode rejection based on block size activity using large block history map (LBHM). Simulation results show that without any meaningful degradation, the proposed method reduces whole encoding time on average by 53%. Also the hybrid usage of the proposed method and the early SKIP mode decision in H.264/AVC reference model reduces whole encoding time by 63% on average.