• Annual Conference on Human and Language Technology
• /
• 2020.10a
• /
• pp.257-260
• /
• 2020

사전 학습을 기반으로 하는 BERT계열의 모델들이 다양한 언어 및 자연어 처리 태스크들에서 뛰어난 성능을 보이고 있지만, masked language model의 경우 입력 문장의 15%만 마스킹을 함으로써 학습 효율이 떨어지고 미세 조정 시 마스킹 토큰이 등장하지 않는 불일치 문제도 존재한다. 이러한 문제를 효과적으로 해결한 ELECTRA는 영어 벤치마크에서 기존의 언어모델들 보다 뛰어난 성능을 보여주었지만 한국어에 대한 관련 연구는 부족한 실정이다. 본 연구에서는 ELECTRA를 한국어 코퍼스에 대해 학습시키고, 다양한 한국어 자연어 이해 태스크들에 대해 실험을 진행한다. 실험을 통해 ELECTRA의 모델 크기별 성능 평가를 진행하였고, 여러 한국어 태스크들에 대해서 평가함으로써 ELECTRA 모델이 기존의 언어 모델들보다 좋은 성능을 보인다는 것을 입증하였다.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2022.07a
• /
• pp.63-66
• /
• 2022

본 논문에서는 분자의 특성을 정확하게 예측하기 위해 효과적인 사전학습(pretraining) 전략과 트랜스포머(Transformer) 모델을 활용한 방법을 제시한다. 딥러닝을 활용한 분자의 성능을 예측하는 연구는 그동안 레이블이 부족한 분자데이터의 특성에 의해 학습 때 사용된 데이터이외의 분자데이터에 대해 일반화 능력이 떨어지는 어려움을 겪었다. 이 논문에서 제시한 모델은 사전학습(pretraining)을 수행할 때 자기지도학습(self-supervised training)을 사용하여 부족한 레이블에 의한 문제점을 피할 수 있다. 대규모 분자 데이터셋으로부터 학습된 이 모델은 4가지 다운스트림 데이터셋에 대해 모두 우수한 성능을 보여주어 일반화 성능이 뛰어나며 효과적인 분자표현을 얻을 수 있음을 보인다.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.644-647
• /
• 2023

대화 시스템에서 대화 상태 추적은 사용자와의 대화를 진행하면서 사용자의 의도를 파악하여 시스템 응답을 결정하는데 있어서 중요한 역할을 수행한다. 특히 목적지향(task-oriented) 대화에서 사용자 목표(goal)를 만족시키기 위해서 대화 상태 추적은 필수적이다. 최근 다양한 자연어처리 다운스트림 태스크들이 사전학습 언어모델을 백본 네트워크로 사용하고 그 위에서 해당 도메인 태스크를 미세조정하는 방식으로 좋은 성능을 내고 있다. 본 논문에서는 한국어 토큰-프리(token-free) 사전학습 언어모델인 KeByT5B 사용하고 종단형(end-to-end) seq2seq 방식으로 미세조정을 수행한 한국어 생성 기반 대화 상태 추적 모델을 소개하고 관련하여 수행한 실험 결과를 설명한다.

• Journal of KIISE:Software and Applications
• /
• v.30 no.7_8
• /
• pp.687-695
• /
• 2003

In this paper, we present a document image compression system based on segmentation, Quincunx downsampling, (5/3) wavelet lifting and subband-oriented SPIHT coding. We reduced the coding time by the adaptation of subband-oriented SPIHT coding and Quincunx downsampling. And to increase compression rate further, we applied arithmetic coding to the bitstream of SPIHT coding output. Finally, we present the reconstructed images for visual comparison and also present the compression rates and PSNR values under various scalar quantization methods.

• Proceedings of the Korean Society of Broadcast Engineers Conference
• /
• 2004.11a
• /
• pp.271-276
• /
• 2004

디지털방송이 시작되면서 고품질의 A/V(비디오/오디오)프로그램과 다양한 멀티미디어 컨텐츠를 제공하는 데이터방송을 처리하기 위한 수신 장치의 비중이 커지고 있다 이와 관련하여 데이터방송 환경에서 새로운 기술과 서비스가 등장할 때마다 이를 수용할 수 있는 소프트웨어를 탑재한 새로운 수신 장치가 필요하다. 일반적으로 한번 가정 내에 보급된 디지털 수신 장치의 소프트웨어 업그레이드가 용이하지 않기 때문에, 방송을 통해 이를 실현한다. 본 논문은 TV 셋탑박스 (STB) 내에 상주하는 middleware native application software 를 방송으로 다운받아 수정하는 기능을 지닌 STB 의 구현에 대해서 기술한다. 소프트웨어 업데이트 시스템은 소프트웨어를 포함하는 데이터 카루셀 스트림을 다운받아 파싱하는 다운로더, 추출된 소프트웨어를 설치하는 업데이트 로더, 그리고 예치상황이 발생하면 셋탑박스가 새로 부팅될 때, 로그 파일을 이용하여 소프트웨어를 옛날 상태의 회복시켜주는 리커버러 (recoverer)로 구성되어 있다. 다운로더는 지상파 디지털 방송 규격인 ATSC 규약에 맞게 구현하고, ATSC용 STB환경에서 테스트하고 있다.

• Electrical & Electronic Materials
• /
• v.9 no.9
• /
• pp.918-924
• /
• 1996

A downstream oxygen plasma is generated by capacitively coupled RF power and applied to photoresist stripping. Stripping rate (ashing rate) is measured in terms of RF power, chamber pressure, oxygen flow rate and temperature. Ashing reaction is thermally activated and depends on oxygen radical density. The ashing process is optimized to have the high ashing rate, good uniformity and minimal plasma damage using a statistical method.

• Journal of Intelligence and Information Systems
• /
• v.28 no.4
• /
• pp.309-328
• /
• 2022

This study empirically analyzed a Korean pre-trained language models (PLMs) designed for natural language generation. The performance of two PLMs - BART and GPT - at the task of abstractive text summarization was compared. To investigate how performance depends on the characteristics of the inference data, ten different document types, containing six types of informational content and creation content, were considered. It was found that BART (which can both generate and understand natural language) performed better than GPT (which can only generate). Upon more detailed examination of the effect of inference data characteristics, the performance of GPT was found to be proportional to the length of the input text. However, even for the longest documents (with optimal GPT performance), BART still out-performed GPT, suggesting that the greatest influence on downstream performance is not the size of the training data or PLMs parameters but the structural suitability of the PLMs for the applied downstream task. The performance of different PLMs was also compared through analyzing parts of speech (POS) shares. BART's performance was inversely related to the proportion of prefixes, adjectives, adverbs and verbs but positively related to that of nouns. This result emphasizes the importance of taking the inference data's characteristics into account when fine-tuning a PLMs for its intended downstream task.

• The Journal of the Korea Contents Association
• /
• v.9 no.8
• /
• pp.73-83
• /
• 2009

TCP(Transmission Control Protocol) is currently used connection-oriented protocol as a typical transport layer protocol in the Internet. However, it has deficiency not be able to communicate with other TCP entities when any link included in the path is down because of single-homing on single path. SCTP(Stream Control Transmission Protocol) suggested as the new transport layer protocol supports multi-homing feature, which provides several paths between source and destination. It can communicate with other SCTP entities using alternate path even when any link on the primary path is down. This paper aims to measure and analyze the multi-homing effect of SCTP over TCP in case of link-down using NS-2 simulator. We classify SCTP into $SCTP_{single-homing}$ and $SCTP_{multi-homing}$ because SCTP with single-homing can also be used like TCP. We measured throughput and bandwidth utilization varying link-down duration, bandwidth, and RTT(round trip time), Simulation results show that throughput of $SCTP_{multi-homing}$ is more than that of TCP by 18 % on average. It is also shown that $SCTP_{multi-homing}$ on varying RTT and bandwidth increases the throughput of TCP by 'l7% and 9% on average, respectively in the link-down environment. In above cases, more or less difference between $SCTP_{single-homing}$ and TCP on throughput and bandwidth utilization was found To summarize, multi-homing effect of SCTP over TCP on throughput is about 18 % on average in the link-down environment This experimental result can be used as the benchmark in order to estimate the multi-homing effect of SCTP over TCP when the link-down happens in the real Internet.

• Journal of Broadcast Engineering
• /
• v.11 no.1 s.30
• /
• pp.66-79
• /
• 2006

As constant development of digital broadcasting and data broadcasting system, new technology will be introduced to digital broadcasting and new broadcasting service will be appeared. These services need to be changed and processed to suit each services and the software of the receiver need to be upgraded. Though, generally the digital broadcasting receivers are not getting updated once it's delivered to home for long time and it need to be done by hand or collected each one of them with trouble. Therefore this paper suggests a way to overcome these difficulties via broadcasting stream. This research is to describe how three modules-namely (1) Downloader, downloads new software from data carrousel stream, (2) Update Loader, installs the software received by downloader; and (3) Recoverer, recovers the former version of the software if some serious problem has been occurred during downloading and installing the software. This paper tries to realize the accommodation of terrestrial STB based on the new technique and service following ATSC A-97 agreement.

• Proceedings of the Korean Institute of Intelligent Systems Conference
• /
• 2007.04a
• /
• pp.465-468
• /
• 2007

This paper addresses the scheme for transmitting Data and TDM signals based on E-PON. E-PON technology, that combines low-cost Ethernet technology and optical fiber infra-structure, has been appeared as a solution of next generation access network. The transmission speed of E-PON is 1Gbps and symmetric in both direction, such as downstream and upstream. Therefore, it is possible to save the cost through sample network architecture, efficient operation, and low maintenance cost of optical IP Ethernet network. By adding TBMoIP(Time Division Multiplexing over Internet Protocol) module to this E-PON system, and implementing QoS(Quality of Service) control function, this system can provide data and TDM service efficiently.