한국어 에세이 문항 자동 채점을 위한 딥러닝 알고리듬 탐색

박강윤,이용상 한국교육평가학회 2022 교육평가연구 Vol.35 No.3

This study was carried out for the purpose of searching for the optimal algorithm for automated scoring system of Korean essay through the comparison of deep learning-based learning models. For this purpose, in this study, deep learning algorithms such as Recurrent Neural Network (RNN), Long-Short-Term-Memory (LSTM), and Gated-Recurrent-Unit (GRU) algorithms were compared. The performance of each algorithm was evaluated based on classification accuracy, precision, recall, and F1. The empirical results showed that the LSTM and GRU algorithm-based models performed better than RNN. Although there is no significant difference in model performance between LSTM and GRU, the GRU algorithm was found to be more efficient in terms of the time required to train the model, so it could be considered to be the optimal algorithm for automated scoring if the machine leanring time is critical.

Yokoyama 술식으로 치료한 근시성 고정사시 1예

박강윤,정재호,최희영,Kang Yeun Pak,MD,Jae Ho Jung,MD,PhD,Hee Young Choi,MD,PhD 대한안과학회 2013 대한안과학회지 Vol.54 No.3

Purpose: To present a surgical result of Yokoyama procedure in myopic strabismus fixus. Case summary: A 51-year-old female patient presented with progressive esotropia and diplopia. According to the Krimsky test, the patient showed 70 prism diopter esotropia and 30 prism diopter hypotropia in her left eye. The axial length was 34.97 mm in the right eye and 33.71 mm in the left eye. The patient was diagnosed with myopic strabismus fixus. The authors performed the Yokoyama procedure on her left eye. Surgical examination revealed each medial rectus muscle was recessed. Half of the muscle bellies of the superior and lateral rectus muscles were sutured together without muscle splitting 15 mm posterior from their insertion. At 1 year postoperatively, the patient showed 30 prism diopter esotropia and 20 prism diopter hypotropia in her right eye by alternative prism cover test. The authors performed the same procedure on her right eye. At 2 months after the second surgery, the patient showed orthotropia in the primary position and gaze limitation was improved. Conclusions: The Yokoyama procedure can be an effective method for improving deviation and ocular motility in esotropia caused by high myopia.

자가결막 피판 전위술을 이용한 군날개 치료의 단기 임상 결과

박강윤,이종수 대한안과학회 2012 대한안과학회지 Vol.53 No.12

Purpose: The present study reviewed the surgical outcomes of conjunctival flap advancement with pterygial tissue excision in primary and recurrent pterygium. Methods: In a retrospective survey of 169 eyes of 169 patients who underwent pterygial tissue excision and conjunctival flap advancement surgery, after a three-month follow-up minimum, history of pterygium surgery, surgical outcomes, recurrence rates, and complications were evaluated. Results: The study included 74 male and 95 female patients. The mean age was 56 ± 11.93 years (range: 31-81 years). One hundred and forty eyes were considered as primary pterygium without history of surgery and 29 eyes were considered as recurrent pterygium. In the cases of primary pterygium, 139 eyes were treated after the surgery without recurrence (success rate: 99.2%) and one recurred case was localized to the sclera. Among the 29 eyes with recurrent pterygium, 26 eyes were treated without recurrence (success rate: 89.7%) and 3 recurred cases were localized to the sclera. There were no significant complications related to the surgical procedures. Conclusions: Conjunctival flap advancement surgery with pterygial tissue excision was an effective method in primary and recurrent pterygium treatment and showed low recurrence and minimal complications.

안내후방콘택트렌즈 삽입술에 의해 발생한 색소녹내장 1예

박강윤,김해송,이지웅.Kang Yeun Pak. MD. Hae Song Kim. MD. PhD. Ji Woong Lee. MD. PhD 대한안과학회 2013 대한안과학회지 Vol.54 No.6

Purpose: To report a case of secondary pigmentary glaucoma after implantable contact lens (ICL) implantation successfully treated with trabeculectomy without ICL removal. Case summary: A 29-year-old woman presented with refractory intraocular pressure (IOP) increase in both eyes. IOP was 22 mm Hg in the right eye and 39 mm Hg in the left eye. The patient received posterior chamber phakic intraocular lens implantation in both eyes 22 months prior. Slit lamp examination revealed patent iridotomy sites in both eyes. Gonioscopy revealed open angles with 4-degree pigment deposits on the trabecular meshwork in both eyes. Ultrasound biomicroscopy examination confirmed contact between ICL and the posterior surface of the iris. In spite of well tolerated medical therapy and selective laser trabeculoplasty, IOP was 46 mm Hg in her left eye. Trabeculectomy was performed in her left eye without ICL removal. At 6 months postoperative, IOP measured 6 mm Hg without any anti-glaucoma medication and bleb was maintained in good condition in the left eye. Conclusions: The results from this case study indicate that ICL implantation can lead to secondary pigmentary glaucoma and trabeculectomy without ICL removal may help to decrease the IOP.

소아 안와연조직염 환자의 임상 특징

박강윤,정재호,최희영,Kang Yeun Pak,Jae Ho Jung,Hee Young Choi 대한안과학회 2013 대한안과학회지 Vol.54 No.3

Purpose: The present study reviews the clinical features of orbital cellulitis in childhood for early diagnosis and proper treatment. Methods: The authors performed a retrospective study by computed tomography (CT) on children under 14 years of age diagnosed with orbital cellulitis and admitted to Pusan National University Hospital from 2003 to 2010. Results: In total, 27 patients were identified (range 4 months to 14 years). Periorbital swelling was the most common initiating symptom, followed by fever and conjunctival injection. Paranasal sinus disease was the most common predisposing factor. Preseptal cellulitis was the most common finding, followed by subperiosteal abscess, orbital cellulitis, and orbital abscess. All patients underwent a blood culture; none were positive. Intravenous antibiotics therapy was performed empirically and was effective in all cases. No patients suffered from permanent complications except recurrence. Conclusions: Orbital cellulitis in children presents with periorbital swelling, fever, conjunctival injection in association with sinusitis, and upper respiratory infection (URI). A CT study is a reliable diagnostic option for the early detection and localization in the pediatric orbital cellulitis. Early empirical antibiotic therapy is mandatory for successful treatment.

순환신경망 장단기 기억(LSTM)을 이용한 자동 채점의 가능성 탐색

박강윤,이용상,신동광 한국교육과정평가원 2021 교육과정평가연구 Vol.24 No.4

본 연구에서는 순환신경망의 일종인 장단기 기억(Long-Short Term Memory: LSTM)를 이용하여 영어 에세이 자동 채점 가능성을 탐색하였다. LSTM은 기존의 순환신경망(Recurrent Neural Network; RNN)이 갖는 장기의존성의 문제를 극복하기 위해 제안된 학습 모델로, 본 연구에서는 이러한 LSTM 모델을 활용하여 영어 에세이 데이터를 학습시킨 후, 별도의 평가 데이터를 통해 LSTM의 성능을 평가하였다. 이분(二分) 자료의 형태를 갖는 선다형 채점 데이터와 달리 에세이 채점 데이터는 다분(多分) 자료의 형태를 가지므로 본 연구에서는 다항 분류가 가능하도록 학습 모델을 구축하여 점수를 예측하였으며, 이러한 LSTM 학습 모델을 여섯 가지 지표(정확성, 정밀도, 재현율, F1, 카파, 상관계수)로 평가하였다. 그 결과를 살펴보면, 본 연구에서 구축한 LSTM 학습 모델이 학생들의 에세이 점수를 양호한 수준에서 예측할 수 있음을 확인하였다. 또한 학습 모델의 성능을 결정하는 주요 요인 중 하나가 데이터의 질과 양임을 감안할 때 향후 충분한 양질의 데이터를 구축하여 학습할 경우 자동 채점의 정확성을 보다 향상시킬 수 있을 것으로 기대된다. 후속 연구로는 최적의 에세이 자동 채점 알고리듬을 도출하기 위해 향후 다양한 순환신경망 모델을 비교 검토하는 실증 연구들이 수행될 필요가 있다. In the present study, the feasibility of an automated essay scoring of English was explored using Long-Short Term Memory (LSTM), a type of Recurrent Neural Network (RNN). LSTM is a deep learning model proposed to overcome the problem of long-term dependence of the existing RNN. In this study, an automated essay scoring model based on LSTM was adopted to score English essay data extracted from the open huge repository of data ‘kaggle,’ and the performance of the model was validated. Unlike multiple-choice scoring data which consisted of binary (true/false) data, essay scoring data had multiple facets, thus the data used for the deep learning model was constructed within a multinomial classification to order to predict scores of those essay data. For its validation, the six indices of ‘accuracy,’ ‘precision,’ ‘recall’, ‘F1-measure,’ ‘kappa,’ and ‘correlation coefficient’ were used. As a result, it turned out that the LSTM model could predict students' essay scores at an appropriate level. The performance of the deep learning model is closely related to the quality and quantity of data, thus it is expected that the accuracy of the automated essay scoring could be improved if sufficient quality data is composed and used for the deep learning process. To derive a more valid and reliable algorithm, it is necessary to conduct further empirical studies by testing various RNN models.

Differences in Clinical Manifestations and Outcomes between Adult and Child Patients with Henoch-Schönlein Purpura

강윤,박진수,강미일,박희진,이상원,이수곤,You-Jung Ha,박용범 대한의학회 2014 Journal of Korean medical science Vol.29 No.2

We aimed to investigate differences in clinical manifestations and outcomes between adult and child patients with Henoch-Schönlein purpura (HSP), and to analyze the factors associated with poor prognosis for HSP nephritis. This retrospective 10-yr study enrolled 160 patients with HSP who visited Severance Hospital. Purpura was mostly detected in lower extremities, but purpura in upper extremities was more frequently observed in adults than children (41.7% vs 19.3%). Children had a greater frequency of arthralgia (55.4% vs 27.1%), while adults had a greater frequency of diarrhea (20% vs 1.6%). Anemia, elevated C-reactive protein, and level of IgA were more frequently observed in adults (25% vs 7.1%, 65.6% vs 38.4%, 26.3% vs 3.5%). Renal involvement in adults was more severe than in children (79.2% vs 30.4%). Chronic renal failure showed a significant difference in outcomes of HSP between adults (10.4%) and children (1.8%) after a follow up period of an average of 27 months. Furthermore, renal insufficiency at diagnosis was significantly related to the progression to chronic renal failure. Our results showed several differences in the clinical features of HSP between adults and children. Adults with HSP had a higher frequency of renal insufficiency and worse renal outcomes than children. Renal insufficiency at diagnosis might be of predictive value for the progression to chronic renal failure in HSP patients.

채점자질을 활용한 인공지능 기반 글쓰기 자동채점 방안 탐색

박종임,최숙기,박강윤,김길재 청람어문교육학회 2023 청람어문교육 Vol.- No.96

This study explored methods to utilize rater characteristics in the automated scoring of Korean writing. To this, we developed and applied 'morpheme-level frequency features' and 'lexical-sentence level deep features‘. The results of this study revealed that the performance of a model using only the morpheme-level frequency features was similar to that of a model incorporating additional deep features. Additionally, the model utilizing the optimal set of scoring features exhibited the highest performance. From these findings, the following implications can be drawn: 1. Basic linguistic features such as morpheme frequency can be effectively used to predict the quality of writing. 2. There is a need to promote research on the development and extraction of deep features at the vocabulary, sentence, and text levels that are effective for analyzing Korean text. 3. Developing writing scoring features requires a comprehensive exploration of various relevant theories, including text analysis theory, writing development theory, rhetoric, and Korean linguistics. 4. Beyond features specialized for specific types of texts or tasks, there is a need to develop generic scoring features that can cover a wide range of text types and tasks. 5. It's essential to explore scoring features for each specific grading category in writing.

청소년 사이버불링 방지를 위한 탐지 모델 개발

박강윤 ( Park Kangyoon ),김길재 ( Kim Giljae ),문명현 ( Moon Myunghyun ) 한국청소년학회 2023 청소년학연구 Vol.30 No.5

본 연구는 청소년 사이버불링 유형 중 사이버 언어폭력을 방지하기 위해 딥러닝을 활용한 언어폭력 탐지 모델을 개발하는 것을 목적으로 한다. 한국지능정보사회진흥원이 AIhub에서 공개한 텍스트 윤리검증 데이터(2022. 10. 갱신)를 사용하였다. 데이터는 총 453,340의 문장이며 윤리 문장과 비윤리 문장은 각각 202,276개, 251,064개이다. 비윤리 문장은 문장의 속성이 욕설(Abuse), 비난(Censure), 범죄(Crime), 차별(Discrimination), 혐오(Hate), 선정(Sexual), 폭력(Violence)으로 총 7개로 라벨링 되어있다. 데이터 불균형을 해결하기 위해 언더샘플링을 통해 전처리를 진행하였다. GRU 알고리즘을 활용하여 딥러닝 탐지 모델을 개발했고 모델 평가 지표를 혼동행렬로 산출했다. 분석결과 Accuracy 69.3%~81.3%, Precision 72%~85.7%, Recall 68.3%~78.7%, F1 Score 70.1%~82.1%의 성능을 보였다. 그 중 Violence로 라벨링된 비윤리 문장 탐지 성능이 가장 높았으며, Hate로 라벨링된 비윤리 문장 탐지 성능이 가장 낮았다. 끝으로 본 연구결과를 토대로 청소년 사이버불링 예방을 위한 논의와 시사점을 제시하였다. The purpose of this study is to develop a language violence detection model using deep learning to prevent cyberbullying, specifically cyber language violence among adolescents. The researchers used the Text Ethical Verification Dataset (updated in October 2022) provided by the NATIONAL INFORMATION SOCIETY AGENCY(AIhub). The dataset contains a total of 453,340 sentences, with 202,276 ethical and 251,064 unethical sentences. The unethical sentences are labeled with one of seven attributes: abuse, censure, crime, discrimination, hate, sexual, and violence. To resolve the data imbalance, the researchers conducted preprocessing through undersampling. Researchers developed a deep learning detection model using the GRU algorithm and evaluated its performance using a confusion matrix as the evaluation metric. The analysis results showed performance ranging from 69.3% to 81.3% for accuracy, 72% to 85.7% for precision, 68.3% to 78.7% for recall, and 70.1% to 82.1% for F1 score. Among them, the detection performance of unethical sentences labeled as violence was the highest, while the performance of those labeled as hate was the lowest. Finally, based on the results of this study, the researchers presented discussions and implications.

커피찌꺼기를 활용하여 제조한 흡착제의 H2S 흡착특성

김정환,강윤성,우제민,박영성 대전대학교 환경문제연구소 2015 환경문제연구소 논문집 Vol.19 No.-

H2S adsorption characteristics of adsorbent made by coffee waste were investigated. For analyses of the manufactured adsorbent, various methods such as scanning electron microscope(SEM), measurements of BET(Brunauer Emmett Teller) surface area, pH, and Iodide adsorption were adopted. As major adsorption characteristics, adsorption equilibrium capacity was measured by using a batch type experimental apparatus for operating variables such as adsorption temperature(25~45℃), adsorbent types. The experimental result showed that the H2S adsorption equilibrium capacity of adsorbent made by coffee waste much more increases with steam activation for the coffee waste.