• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.61-65
• /
• 2023

최근, 대부분의 태스크가 초거대 언어 모델로 통합되고 있을 정도로 많은 관심 및 연구되고 있다. 초거대 언어 모델을 효과적으로 활용하기 위해서는 모델의 능력에 대한 분석이 선행되어야 하나, 한국어에 대한 분석 및 탐색은 상대적으로 부족하다. 본 논문에서는 한국어 맞춤법 교정 태스크를 통해 초거대 언어 모델의 능력을 탐색한다. 맞춤법 교정 태스크는 문장의 구조 및 문법을 이해하는 능력이 필요하며, 사용자의 만족도에 영향을 미칠 수 있는 중요한 태스크이다. 우리는 맞춤법 세부 유형에 따른 ChatGPT의 제로샷 및 퓨샷성능을 평가하여 초거대 언어 모델의 성능 분석을 수행한다. 실험 결과 제로샷의 경우 문장부호 오류의 성능이 가장 우수했으며, 수사 오류의 성능이 가장 낮았다. 또한, 예제를 더 많이 제공할수록 전체적인 모델의 성능이 향상되었으나, 제로샷의 경우보다 오류 유형 간의 성능 차이가 커지는 것을 관찰할 수 있었다.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.726-729
• /
• 2023

본 논문에서는 대화 중 발화에서 페르소나 트리플을 추출하는 방법을 연구한다. 발화 문장과 그에 해당하는 트리플 쌍을 활용하여 발화 문장 혹은 페르소나 문장이 주어졌을 때 그로부터 페르소나 트리플을 추출하도록 모델을 멀티 태스크 러닝 방식으로 학습시킨다. 모델은 인코더-디코더 구조를 갖는 사전학습 언어모델 BART [1]와 T5 [2]를 활용하며 relation 추출과 tail 추출의 두 가지 태스크를 각각 인코더, 디코더 위에 head를 추가하여 학습한다. Relation 추출은 분류로, tail 추출은 생성 문제로 접근하도록 하여 최종적으로 head, relation, tail의 구조를 갖는 페르소나 트리플을 추출하도록 한다. 실험에서는 BART와 T5를 활용하여 각 태스크에 대해 다른 학습 가중치를 두어 훈련시켰고, 두 모델 모두 relation과 tail을 추출하는 태스크 정확도에 있어서 90% 이상의 높은 점수를 보임을 확인했다.

• Journal of KIISE:Software and Applications
• /
• v.37 no.1
• /
• pp.19-28
• /
• 2010

Recently, various home appliances are inter connected in wired/wireless network to provide a single service, but their interfaces are not uniformly integrated. In order to get a service in such an environment, users have to figure out how to control each appliance and the way of combining its functions with others. Even worse to the users, many buttons are barely used to control the complicated functions of the appliances on the remote controllers, and controllers have very similar shapes in each other. In this paper, based on the definitions of a task as a unit of services provided in the environment of multi-connected appliances, we propose a markup language for home appliances, called HAML (Home Appliance Markup Language), for generating task-based user interfaces to help controlling multi-connected appliances as one pleases. The proposed method generates interfaces by gathering the buttons frequently used and necessary for tasks, and provides automatic settings of multi-connected appliances. The proposed method is verified with an analysis of scenarios and usability tests.

• The Journal of the Korea Contents Association
• /
• v.18 no.3
• /
• pp.101-113
• /
• 2018

Systems such as military inspection equipment which it is important to acquire and evaluate data in real-time should be able to real-time processing at the operating system level. As technology advances, there is a demand for replacing existing equipment with mobile device, but mobile devices with Android are not suitable for systems requiring real-time performance. On Android, garbage collection ensures free memory, while other tasks are interrupted while this task is performed, which cannot guarantee periodicity of particular tasks. In this paper, we designed and implemented a structure to control execution garbage collection of Android to solve this problem. Real-time performance is ensured by controlling garbage collection during the time required for real-time operation, and RTiK(Real-Time implanted Kernel) is applied to ensure real-time performance on Android. In order to evaluate the performance, we measured the call period of the 5ms period task, and, only 34.31% of the task was guaranteed before the control, but the task period of 98.18% was satisfied through control, providing real-time performance to Android.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.16 no.10
• /
• pp.2303-2308
• /
• 2012

In this paper, we deal with the problem of scheduling parallel tasks with deadlines. Parallel tasks can be simultaneously executed on various machines and specially, we consider the malleable tasks, that is, the tasks whose execution time is given by a function of the number of machines on which they are executed. The goal of the problem is to maximize the throughput of tasks completed within their deadlines. This problem is well-known as NP-hard problem. Thus we will find an approximation algorithm, and its performance is compared with that of the optimal algorithm and analyzed by finding the approximation ratio. In particular, the algorithm has more resources, that is, more machines, than the optimal algorithm. This is called the resource augmentation analysis. We propose an algorithm to guarantee the approximation ratio of 3.67 using 1.5 times machines.

• Journal of Internet Computing and Services
• /
• v.18 no.1
• /
• pp.1-9
• /
• 2017

Innovations from current researches on cloud computing such as applying bio-inspired computing techniques have brought new level solutions in offloading mechanisms. With the growing trend of mobile devices, mobile cloud computing can also benefit from applying bio-inspired techniques. Energy-efficient offloading mechanisms on mobile cloud systems are needed to reduce the total energy consumption but previous works did not consider energy consumption in the decision-making of task distribution. This paper proposes the Particle Swarm Optimization (PSO) as an offloading strategy of cloudlet to data centers where each task is represented as a particle during the process. The collected tasks are classified using K-means clustering on the cloudlet before applying PSO in order to minimize the number of particles and to locate the best data center for a specific task, instead of considering all tasks during the PSO process. Simulation results show that the proposed PSO excels in choosing data centers with respect to energy consumption, while it has accumulated a little more processing time compared to the other approaches.

• Journal of Korea Multimedia Society
• /
• v.16 no.4
• /
• pp.493-501
• /
• 2013

A key challenge in massively multiplayer on-line game(MMOG) systems is providing real-time response latencies to the large number of concurrent game players. MMOG systems are a kind of soft real-time systems because requests from many players should be responded within specified time constraints. Client events have different timeliness and consistency requirements according to their nature in the game world. These requirements lead to tasks with different priorities on CPU processing. In order to meet their timing constraints, we propose priority scheduling methods that attempt to allocate preferentially more CPU bandwidth to serve an task with the higher priority level in the presence of transient overloading. The proposed scheduling methods are capable of enhancing real-time performance of MMOG system by maximizing the number of tasks with higher priority completed successfully within their deadlines while minimizing total average latency of tasks finished after given deadlines. The performance of these scheduling methods is evaluated through extensive simulation experiments.

• Journal of Korea Multimedia Society
• /
• v.13 no.10
• /
• pp.1423-1435
• /
• 2010

Building wireless sensor networks requires a constituting sensor node to consider the following limited hardware resources: a small battery lifetime limiting available power supply for the sensor node, a low-power microprocessor with a low-performance computing capability, and scarce memory resources. Despite such limited hardware resources of the sensor node, the sensor node platform needs to activate real-time sensing, guarantee the real-time processing of sensing data, and exchange data between individual sensor nodes concurrently. Therefore, in this paper, we propose an energy-efficient real-time task scheduling technique for battery-powered wireless sensor nodes. The proposed energy-efficient task scheduling technique controls the microprocessor's operating frequency and reduces the power consumption of a task by exploiting the slack time of the task when the actual execution time of the task can be less than its worst case execution time. The outcomes from experiments showed that the proposed scheduling technique yielded efficient performance in terms of guaranteeing the completion of real-time tasks within their deadlines and aiming to provide low power consumption.

• Journal of KIISE:Computer Systems and Theory
• /
• v.35 no.6
• /
• pp.273-283
• /
• 2008

Many embedded real-time systems have adopted processors supported with dynamic voltage scaling(DVS) recently. Power is one of the important metrics for optimization in the design and operation of embedded real-time systems. We can save considerable energy by using slowdown of processor supported with DVS. In this paper, we propose heuristic algorithms to calculate task slowdown factors for an efficient energy consumption in embedded real-time systems with task synchronization. The previous algorithm has a following constraint : given the tasks are ordered in a nondecreasing order of their relative deadline, the task slowdown factors computed are in a nonincreasing order. In this paper, we relax the constraint and propose heuristic algorithms which have the same time complexity that previous algorithm has and can save more energy. Experimental results show that the proposed algorithms are energy efficient.

• Journal of KIISE
• /
• v.43 no.8
• /
• pp.893-901
• /
• 2016

In Internet of Things (IoT) environments, users are required to search for IoT devices necessary to access services for accomplishing their tasks. As IoT technologies advance, a user task will utilize various types of IoT-based services that are deployed in an IoT environment. Therefore, to accomplish a user task effectively, the services that utilize IoT devices need to be found in a certain geographical region. In addition, the service discovery needs to be accomplished in a stable manner while considering dynamically changing IoT environments. To deal with these issues, we propose two service discovery methods that consider geographic cohesiveness of services in IoT environments. We compare the effectiveness of the proposed methods against a traditional service discovery algorithm that does not consider geographic cohesiveness.