• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.05a
• /
• pp.684-686
• /
• 2023

최근 대규모 언어모델 (large language models) 을 활용하여 다양한 자연어처리 문제를 추가학습 없이 풀어내기 위한 zero-shot 학습에 대한 연구가 활발히 수행되고 있다. 특히 프롬프트 튜닝(prompt tuning)을 활용하여 적은 학습만으로도 효과적으로 다양한 태스크에 적응하도록 돕는 방법이 최근 대규모 언어모델의 성능을 향상시키고 있다. 본 논문은 명령어 튜닝 (instruction tuning) 이 언어모델에 끼치는 영향을 분석하였다. 명령어 튜닝된 모델이 기존 언어모델과 비교하여 변화된 문장 생성 특징, 생성된 문장의 품질 등에 대한 분석을 수행하고 결과를 제시한다.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.437-442
• /
• 2023

최근에 등장한 대규모 언어 모델은 다양한 언어 처리 작업에서 놀라운 성능을 발휘하고 있다. 그러나 이러한 모델의 크기와 복잡성 때문에 모델 경량화의 필요성이 대두되고 있다. Pruning은 이러한 경량화 전략 중 하나로, 모델의 가중치나 연결의 일부를 제거하여 크기를 줄이면서도 동시에 성능을 최적화하는 방법을 제시한다. 본 논문에서는 한국어 언어 모델인 Polyglot-Ko에 Wanda[1] 기법을 적용하여 Pruning 작업을 수행하였다. 그리고 이를 통해 가중치가 제거된 모델의 Perplexity, Zero-shot 성능, 그리고 Fine-tuning 후의 성능을 분석하였다. 실험 결과, Wanda-50%, 4:8 Sparsity 패턴, 2:4 Sparsity 패턴의 순서로 높은 성능을 나타냈으며, 특히 일부 조건에서는 기존의 Dense 모델보다 더 뛰어난 성능을 보였다. 이러한 결과는 오늘날 대규모 언어 모델 중심의 연구에서 Pruning 기법의 효과와 그 중요성을 재확인하는 계기가 되었다.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2024.05a
• /
• pp.509-512
• /
• 2024

자연어 처리 기술은 인간 언어의 이해와 처리에서 큰 진전을 이루었으나, 학습 데이터에 내재한 성별 편향이 모델의 예측 정확도와 신뢰성을 저하하는 주요한 문제로 남아 있다. 특히 성별 예측에서 이러한 편향은 더욱 두드러진다. 제로샷 분류 기법은 기존에 학습되지 않은 새로운 클래스를 효과적으로 예측할 수 있는 기술로, 학습 데이터의 제한적인 의존성을 극복하고 다양한 언어 및 데이터 제한 상황에서도 효율적으로 작동한다. 본 논문은 성별 클래스 확장과 데이터 구조 개선을 통해 성별 편향을 최소화한 새로운 데이터셋을 구축하고, 이를 제로샷 분류 기법을 통해 학습시켜 성별 편향성이 완화된 새로운 성별 예측 모델을 제안한다. 이 연구는 다양한 언어로 구성된 자연어 데이터를 추가 학습하여 성별 예측에 최적화된 모델을 개발하고, 제한된 데이터 환경에서도 모델의 유연성과 범용성을 입증한다.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2024.05a
• /
• pp.570-573
• /
• 2024

본 연구는 대형 언어 모델 (LLM) 시대에 공감적 대화 생성을 위한 감정 인식의 필요성을 확인하고 소형 언어 모델 (SLM)을 통한 미세 조정 학습이 고비용 LLM, 특히 ChatGPT의 대안이 될 수 있는지를 탐구한다. 이를 위해 KoBERT 미세 조정 모델과 ChatGPT를 사용하여 사용자 감정을 인식하고, Polyglot-Ko 미세 조정 모델 및 ChatGPT를 활용하여 공감적 응답을 생성하는 비교 실험을 진행하였다. 실험 결과, KoBERT 기반의 감정 분류기는 ChatGPT의 zero-shot 접근 방식보다 뛰어난 성능을 보였으며, 정확한 감정 분류가 공감적 대화의 질을 개선하는 데 기여함을 확인하였다. 이는 공감적 대화 생성을 위해 감정 인식이 여전히 필요하며, SLM의 미세 조정이 고비용 LLM의 실용적 대체 수단이 될 수 있음을 시사한다.

• The Journal of Korea Robotics Society
• /
• v.19 no.1
• /
• pp.16-22
• /
• 2024

Robotic grasping in unstructured environments poses a significant challenge, demanding precise estimation of gripping positions for diverse and unknown objects. Generative Grasping Convolution Neural Network (GG-CNN) can estimate the position and direction that can be gripped by a robot gripper for an unknown object based on a three-dimensional depth map. Since GG-CNN uses only a depth map as an input, the precision of the depth map is the most critical factor affecting the result. To address the challenge of depth map precision, we integrate the Segment Anything Model renowned for its robust zero-shot performance across various segmentation tasks. We adjust the components corresponding to the segmented areas in the depth map aligned through external calibration. The proposed method was validated on the Cornell dataset and SurgicalKit dataset. Quantitative analysis compared to existing methods showed a 49.8% improvement with the dataset including surgical instruments. The results highlight the practical importance of our approach, especially in scenarios involving thin and metallic objects.

• Cartoon and Animation Studies
• /
• s.34
• /
• pp.105-128
• /
• 2014

This study was aimed to analyze the disparities of 3D stereoscopic images in four well-known American animation movies. After Avatar (2009), lots of stereoscopic movies were developed in Korean 3D production. Almost all 3D productions in Korea, however, focus on the display images or TV series animation yet. In order to make many well-made Korean stereoscopic 3D animations in future, analyzing and comparing the disparities of 3D stereoscopic images is necessary and even mandated. First, I chose 40 cuts from each four American stereoscopic 3D feature films, including Despicable me 2, Epic, Monster University, and Turbo. According to the classifications of shot angles by Vineyard (2008), secondly I analyze the 23 different angular disparities of 3D stereoscopic images and displayed in tables. Demonstrated shot angle disparities in each scene would provide numerical information to animators how to design and make the 3D stereoscopic images. Making successful stereoscopic 3D feature film will be a huge turning point in the Korean animation field in future. This study would be a first trial to seek a new method to set ahead an outlook of numerical values of 3D stereoscopic images for better visual effects.

• 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.

• Journal of the Korea Society of Computer and Information
• /
• v.28 no.8
• /
• pp.195-204
• /
• 2023

In this study, prompt engineering that improves prompts was explored to improve the performance of teaching and learning materials recommendations using generative artificial intelligence such as GPT and Stable Diffusion. Picture materials were used as the types of teaching and learning materials. To explore the impact of the prompt composition, a Zero-Shot prompt, a prompt containing learning target grade information, a prompt containing learning goals, and a prompt containing both learning target grades and learning goals were designed to collect responses. The collected responses were embedded using Sentence Transformers, dimensionalized to t-SNE, and visualized, and then the relationship between prompts and responses was explored. In addition, each response was clustered using the k-means clustering algorithm, then the adjacent value of the widest cluster was selected as a representative value, imaged using Stable Diffusion, and evaluated by 30 elementary school teachers according to the criteria for evaluating teaching and learning materials. Thirty teachers judged that three of the four picture materials recommended were of educational value, and two of them could be used for actual classes. The prompt that recommended the most valuable picture material appeared as a prompt containing both the target grade and the learning goal.

• Transactions of the Korean Society of Mechanical Engineers
• /
• v.12 no.3
• /
• pp.534-540
• /
• 1988

A study on the mixing process of fuel with ambient gas is necessary to verify combustion process of a diesel engine, especially the mechanism of its ignition delay. In this study, a single shot of diesel spray was injected through either a constant pressure injection system and bypass type injection system. Measurements were made on the flow characteristics of ambient gas and its time history using a hot wire anemometer and a high speed camera. The gas flow direction was determined by a smoke tracer method. (1) The ambient gas of spray flows away at the stagnation part where static pressure value is positive and flows in at the penetration part of a negative value with the steady entrainment length of 0.7. (2) The steady entertainment velocity around the spray in creases from the nozzle tip to the downstream, has the maximum value at the mixing boundary part, and represents zero at the stagnation boundary part after which the stream flows reversely at the stagnation part.

• Journal of Internet Computing and Services
• /
• v.23 no.2
• /
• pp.45-52
• /
• 2022

With the development of artificial intelligence analysis methods, especially machine learning, various fields are widely expanding their application ranges. However, in the case of classical music, there still remain some difficulties in applying machine learning techniques. Genre classification or music recommendation systems generated by deep learning algorithms are actively used in general music, but not in classical music. In this paper, we attempted to classify opera among classical music. To this end, an experiment was conducted to determine which criteria are most suitable among, composer, period of composition, and emotional atmosphere, which are the basic features of music. To generate emotional labels, we adopted zero-shot classification with four basic emotions, 'happiness', 'sadness', 'anger', and 'fear.' After embedding the opera libretto with the doc2vec processing model, the optimal number of clusters is computed based on the result of the elbow method. Decided four centroids are then adopted in k-means clustering to classify unsupervised libretto datasets. We were able to get optimized clustering based on the result of adjusted rand index scores. With these results, we compared them with notated variables of music. As a result, it was confirmed that the four clusterings calculated by machine after training were most similar to the grouping result by period. Additionally, we were able to verify that the emotional similarity between composer and period did not appear significantly. At the end of the study, by knowing the period is the right criteria, we hope that it makes easier for music listeners to find music that suits their tastes.