• ETRI Journal
• /
• v.42 no.6
• /
• pp.899-911
• /
• 2020

The data in tables are accurate and rich in information, which facilitates the performance of information extraction and question answering (QA) tasks. TableQA, which is based on tables, solves problems by understanding the table structure and searching for answers to questions. In this paper, we introduce both novice and intermediate Korean TableQA tasks that involve deducing the answer to a question from structured tabular data and using it to build a question answering pair. To solve Korean TableQA tasks, we use S3-NET, which has shown a good performance in machine reading comprehension (MRC), and propose a method of converting structured tabular data into a record format suitable for MRC. Our experimental results show that the proposed method outperforms a baseline in both the novice task (exact match (EM) 96.48% and F1 97.06%) and intermediate task (EM 99.30% and F1 99.55%).

• Annual Conference on Human and Language Technology
• /
• 2018.10a
• /
• pp.273-277
• /
• 2018

기계가 주어진 텍스트를 이해하고 추론하는 능력을 기계독해 능력이라 한다. 기계독해는 질의응답 태스크에 적용될 수 있는데 이것을 기계독해 질의응답이라 한다. 기계독해 질의응답은 주어진 질문과 문서를 이해하고 이를 기반으로 질문에 적합한 답을 출력하는 태스크이다. 본 논문에서는 구조화된 표 형식 데이터로부터 질문에 대한 답을 추론하는 TableQA 태스크를 소개하고, $S^3-NET$을 이용하여 TableQA 문제를 해결할 것을 제안한다. 실험 결과, 본 논문에서 제안한 방법이 EM 96.36%, F1 97.04%로 우수한 성능을 보였다.

• Annual Conference on Human and Language Technology
• /
• 2018.10a
• /
• pp.263-269
• /
• 2018

본 논문에서는 실생활에서 쓰이는 다양한 구조를 갖는 문서에 대해서도 자연어 질의응답이 가능한 모델을 만들고자, 그 첫걸음으로 표에 대해 자연어 질의응답이 가능한 End-to-End 인공신경망 모델 TabQA를 제안한다. TabQA는 기존 연구들과는 달리 표의 형식에 구애받지 않고 여러 가지 형태의 표를 처리할 수 있으며, 다양한 정보의 인코딩으로 풍부해진 셀의 feature를 통해, 표의 row와 column 객체를 직관적이고도 효과적으로 추상화한다. 우리는 본 연구의 결과를 검증하기 위해 다채로운 어휘를 가지는 표 데이터에 대한 질의응답 쌍을 자체적으로 생성하였으며, 이에 대해 단일 모델 EM 스코어 96.0%에 이르는 결과를 얻었다. 이로써 우리는 추후 더 넓은 범위의 양식이 있는 데이터에 대해서도 자연어로 질의응답 할 수 있는 가능성을 확인하였다.

• Annual Conference on Human and Language Technology
• /
• 2023.10a
• /
• pp.75-79
• /
• 2023

문서에서 테이블은 중요한 정보들을 축약하여 모아 놓은 정보 집합체라고 할 수 있다. 이러한 테이블을 대상으로 질의응답하는 테이블 질의응답 기술이 연구되고 있으며, 이 중 언어모델을 이용한 연구가 좋은 결과를 보이고 있다. 본 연구에서는 최근 주목받고 있는 생성형 언어모델 기술을 테이블 질의응답에 적용하여 언어모델과 프롬프트의 변경에 따른 결과를 살펴보고, 단답형 정답과 생성형 결과의 특성에 적합한 평가방법으로 측정해 보았다. 자체 개발한 EXAONE 1.7B 모델의 경우 KorWiki 데이터셋에 대해 적용하여 EM 92.49, F1 94.81의 결과를 얻었으며, 이를 통해 작은 크기의 모델을 파인튜닝하여 GPT-4와 같은 초거대 모델보다 좋은 성능을 보일 수 있음을 확인하였다.

• The Journal of Korean Society for Radiation Therapy
• /
• v.21 no.2
• /
• pp.89-95
• /
• 2009

Purpose: The purpose is to evaluate efficiency of the CT simulator QA phantom manufactured for daily QA. Materials and Methods: We made holes ($1{\times}100{\times}1\;mm$) to verify accuracy between image and real measurement in polystyrene phantom and made 1 mm holes to verify table movement accuracy at superior and inferior 100 mm to the center of the phantom and inserted radiopacity material. To evaluate laser alignment, we made cross mark on the right and left side at phantom and to evaluate CT number accuracy we made 3 cylindrical holes and inserted equivalence material of bone, water, air in them. After CT scanning the phantom, We evaluated accuracy between image and real measurement, accuracy of table movement, laser, and CT number using exposed image. Results: It was measured that the accuracy between image and real measurement was ${\pm}0.3\;mm$, table movement accuracy was ${\pm}0.3\;mm$, laser accuracy was ${\pm}0.5\;mm$ from 7th January to 7th March in 2008 as within the reference point ${\pm}1\;mm$. In the CT number accuracy of bone was ${\pm}10\;HU$, air was ${\pm}5\;HU$, water was ${\pm}5\;HU$ as within the reference point is ${\pm}10\;HU$. Conclusion: We was able to perform CT simulator QA and laser equipment QA more conveniently and fast using manufactured phantom at the same time. We will be able to make more accurate treatment plan that added to QA procedures using images at previous daily QA.

• Progress in Medical Physics
• /
• v.15 no.3
• /
• pp.161-166
• /
• 2004

A simplistic quality assurance (QA) method was designed for a Linac built-in enhanced dynamic wedge (EDW), which can be utilized to make wedged beam distributions. For the purpose of implementing the EDW symmetry QA, a film dosimetry system, low speedy dosimetry film, film densitometer and 3D RTP system were used, and the films irradiated by means of a 60$^{\circ}$ Reversed wedge pair (REWP) method. The profiles were then analyzed in terms of their symmetries, including partial treatment, which is the case of stopping it abruptly during EDW irradiation, and the measured and calculated values compared using the Cad Plan Golden Segmented Treatment Table (Golden STT). The result of this experiment was in good agreement, within 1 %, of the 'reversed wedge pair counterbalance effect'. For the QA of the effective wedge factor (EWF), the authors measured EWFs in relation to the 10$^{\circ}$, 15$^{\circ}$, 20$^{\circ}$, 25$^{\circ}$, 30$^{\circ}$, 45$^{\circ}$ and 60$^{\circ}$ EDW, which were compared with the calculated values using the correction factor derived from the Golden STT and the log files produced automatically during the process of EDW irradiation. By means of this method it was capable of check up the safety of effective wedge factor without any other dosimetry system. The EDW QA was able to be completed within 1 hour from irradiation to analysis as a consequence of the simplified QA procedure, with maximized effectiveness. Unlike the metal wedge system, the EDW system was heavily dependent on the dose rates and jaw movements; therefore, its features could potentially cause inaccuracy. The frequent simplistic QA for the EDW is essential, and could secure against the flaw of dynamic treatment that uses the EDW.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2022.10a
• /
• pp.164-167
• /
• 2022

In table Machine comprehension, the knowledge required for language models or the structural form of tables changes depending on the domain, showing a greater performance degradation compared to text data. In this paper, we propose a pre-learning data construction method and an adversarial learning method through meaningful tabular data selection for constructing a pre-learning table language model robust to these domain changes in table machine reading. In order to detect tabular data sed for decoration of web documents without structural information from the extracted table data, a rule through heuristic was defined to identify head data and select table data was applied. An adversarial learning method between tabular data and infobax data with knowledge information about entities was applied. When the data was refined compared to when it was trained with the existing unrefined data, F1 3.45 and EM 4.14 increased in the KorQuAD table data, and F1 19.38, EM 4.22 compared to when the data was not refined in the Spec table QA data showed increased performance.

• The Journal of Korean Society for Radiation Therapy
• /
• v.11 no.1
• /
• pp.16-21
• /
• 1999

Purpose : The aim of this study is to conform the possibility of the liquid type EPID as a QC tools to clinical indication and of replacement of the film dosimetry. Aditional aim is to describe a procedure for the use of a EPID as a physics calibration tool in the measurements of radiation beam parameters which are typically carried out with film. Method & Materials : In this study we used the Clinac 2100c/d with EPID. This system contains 65536 liquid-filled ion chambers arranged in a $256{\times}256$ matrix and the imaging area is $32.5{\times}32.5cm$ with liquid layer thickness of 1mm. The EPID was tested for different field sizes under typical clinical conditions and pixel values were calibrated against dose by producing images using various thickness of lead attenuators(lead step wedge) using 6 & 10MV x-ray. We placed various thickness of lead on the table of linear accelerator and set the portal vision an SDD of 100cm. To acquire portal image we change the field size and energy, and we recorded the average pixel value in a $3{\times}3$ pixel region of interest(ROI) at field center was recorded. The pixel values were also measured for different field sizes in order to evaluate the dependence of pixel value on x-ray energy spectrum and various scatter components. Result : The EPID, as a whole, was useful as a QA tool and dosimetry device. In mechanical check, cross-hair centering was well matched and the error was less than ?2mm and light/radiation field coincidence was less than 1mm also. In portal dosimetry the wider the field size the the higher the pixel value and as the lead thickness increase, the pixel value was exponentially decreased. Conclusions : The EPID was very suitable for QA tools and it can be used to measure exit dose during patients treatment with reasonable accuracy. But when indicate the EPID to clincal study deep consideration required

• The Journal of Korean Society for Radiation Therapy
• /
• v.17 no.1
• /
• pp.41-43
• /
• 2005

Purpose : Wish to present degree management process that is efficient confirm radiation treatment exclusive use CT simulator's Q.A item that become Q.A and Differentiation of diagnosis area that present Report of the AAPM Task Group No.66 using Q.A tool that produce itself and secure safe and correct CT-simulation process and equip convenience. Method and material : Manufacture CT simulator's Q.A tool on source and confirm virtue between isocenter of wall laser system, patient table, CT scanner's imaging plane that present in Report of the AAPM Task Group No.66 by daily publication unit. Result : Confirmed measured value from Report of the AAPM Task Group No.66 to confirmation of presenting degree management item in wall laser's ${\pm}2mm$, table's ${\pm}2mm$, imaging plane's ${\pm}2mm$ tolerance extent. Conclusion : There is unconfirmed item from CT-simulation process for therapy to CT Q.A protocol of existent diagnosis area, premising suitable degree management of radiation treatment exclusive use CT-simulator equipment confirming presenting Q.A item in Report of the AAPM Task Group No.66 safe and correct CT-simulation process guarantee can

• The Journal of the Korea Contents Association
• /
• v.9 no.11
• /
• pp.254-261
• /
• 2009

This study suggested that the table of CT-simulator and the laser alignment system using diagnostic CT scanner have an efficient method for improvement in alignment between the planned target center of traverse image with CT scanner. It was conducted on the daily QA when presented in the AAPM TG66 with correcting the laser alignment system using geometric trigonometric functions and investigated the effectiveness of correction methods as compared with those before and after correction. Before correction error was 3.82mm between the planned target center of image, the table longitudinal axis was twisted with 0.436o. The laser alignment system using geometric trigonometric functions in after correction was satisfied with tolerance limits of ${\pm}2mm$ when occurred about 0.7mm in errors between the planned target center. The table correction to satisfy the geometric accuracy is very inefficient over against the time and economic loss as well as technical limits in the case of application as only radiation therapy associated with CT-simulator with diagnostic CT scanner in use. But, the method which corrects the laser alignment system is economic and relatively simple with possibility of getting well geometric accuracy and we suppose that it is efficient method for applying in the clinic.