• 한국융합학회논문지
• /
• 제13권3호
• /
• pp.77-89
• /
• 2022

대용량의 코퍼스로 학습한 사전학습 언어모델이 다양한 자연어처리 태스크에서 성능 향상에 도움을 주는 것은 많은 연구를 통해 증명되었다. 하지만 자원이 부족한 언어 환경에서 사전학습 언어모델 학습을 위한 대용량의 코퍼스를 구축하는데는 한계가 있다. 이러한 한계를 극복할 수 있는 Cross-lingual Post-Training (XPT) 방법론을 사용하여 비교적 자원이 부족한 한국어에서 해당 방법론의 효율성을 분석한다. XPT 방법론은 자원이 풍부한 영어의 사전학습 언어모델의 파라미터를 필요에 따라 선택적으로 재활용하여 사용하며 두 언어 사이의 관계를 학습하기 위해 적응계층을 사용한다. 이를 통해 관계추출 태스크에서 적은 양의 목표 언어 데이터셋만으로도 원시언어의 사전학습 모델보다 우수한 성능을 보이는 것을 확인한다. 더불어, 국내외 학계와 기업에서 공개한 한국어 사전학습 언어모델 및 한국어 multilingual 사전학습 모델에 대한 조사를 통해 각 모델의 특징을 분석한다

• 정보처리학회논문지:소프트웨어 및 데이터공학
• /
• 제12권8호
• /
• pp.341-354
• /
• 2023

문서의 제목은 문서의 내용을 가장 효율적으로 요약하여 제공해 준다. 이때 독자들이 선호하는 스타일과 언어에 따라 문서의 제목을 다르게 제공해 준다면, 독자들은 문서의 내용을 좀 더 쉽게 예측할 수 있다. 본 연구에서는 문서가 주어졌을 때 언어와 스타일에 따라 제목을 자동 생성하는'교차 언어 및 스타일 기반의 제목 생성 모델을 제안한다. 모델을 학습하기 위해서는 같은 내용을 다른 언어와 다른 스타일로 작성한 병렬데이터가 필요하다. 그러나 이러한 종류의 병렬데이터는 구축하기 매우 어렵다. 반면, 단일 언어와 단일 스타일로 구축된 제목 생성 데이터는 많으므로 본 연구에서는 제로샷(zero-shot) 학습으로 제목 생성을 수행하고자 한다. 교차 언어 및 스타일 기반의 제목 생성을 학습하기 위해 다중 언어로 사전 학습된 트랜스포머 모델에 각 언어, 스타일, 기계번역을 위한 어댑터를 추가하였다. 기계 번역용 병렬데이터를 이용하여 기계번역을 먼저 학습한 후, 동일 스타일의 제목 생성을 학습하였다. 이때, 필요한 어댑터만을 학습하고 다른 부분의 파라미터는 모두 고정시킨다. 교차 언어 및 스타일 기반의 제목을 생성할 때에는 목적 언어와 목적 스타일에 해당하는 어댑터만을 활성화시킨다. 실험 결과로는 각 모델을 따로 학습시켜 파이프라인으로 연결시킨 베이스라인에 비해 본 연구에서 제안한 제로샷 제목 생성의 성능이 크게 떨어지지 않았다. 최근 대규모 언어 모델의 등장으로 인한 자연어 생성에서의 많은 변화가 있다. 그러나 제한된 자원과 제한된 데이터만을 이용하여 자연어 생성의 성능을 개선하는 연구는 계속되어야 하며, 그런 점에서 본 연구의 의의를 모색한다.

• Journal of Periodontal and Implant Science
• /
• 제50권1호
• /
• pp.28-37
• /
• 2020

Purpose: The aim of our study was to determine the prevalence and degree of lingual concavities in the first molar region of the mandible to reduce the risk of perforating the lingual cortical bone during dental implant insertion. Methods: A total of 163 suitable cross-sectional cone-beam computed tomography images of edentulous mandibular first molar regions were evaluated. The mandibular morphology was classified as a U-configuration (undercut), a P-configuration (parallel), or a C-configuration (convex), depending on the shape of the alveolar ridge. The characteristics of lingual concavities, including their depth, angle, vertical location, and additional parameters, were measured. Results: Lingual undercuts had a prevalence of 32.5% in the first molar region. The mean concavity angle was 63.34°±8.26°, and the mean linear concavity depth (LCD) was 3.03±0.99 mm. The mean vertical distances of point P from the alveolar crest (Vc) and from the inferior mandibular border were 9.39±3.39 and 16.25±2.44, respectively. Men displayed a larger vertical height from the alveolar crest to 2 mm coronal to the inferior alveolar nerve (Vcb) and a wider LCD than women (P<0.05). Negative correlations were found between age and buccolingual width at 2 mm apical to the alveolar crest, between age and Vcb, between age and Vc, and between age and LCD (P<0.05). Conclusions: The prevalence of lingual concavities was 32.5% in this study. Age and gender had statistically significant effects on the lingual morphology. The risk of lingual perforation was higher in young men than in the other groups analyzed.

• Journal of Information Processing Systems
• /
• 제13권4호
• /
• pp.928-940
• /
• 2017

State-of-the-art speaker recognition systems may work better for the English language. However, if the same system is used for recognizing those who speak different languages, the systems may yield a poor performance. In this work, the decisions of a Gaussian mixture model-universal background model (GMM-UBM) and a learning vector quantization (LVQ) are combined to improve the recognition performance of a multilingual speaker identification system. The difference between these classifiers is in their modeling techniques. The former one is based on probabilistic approach and the latter one is based on the fine-tuning of neurons. Since the approaches are different, each modeling technique identifies different sets of speakers for the same database set. Therefore, the decisions of the classifiers may be used to improve the performance. In this study, multitaper mel-frequency cepstral coefficients (MFCCs) are used as the features and the monolingual and cross-lingual speaker identification studies are conducted using NIST-2003 and our own database. The experimental results show that the combined system improves the performance by nearly 10% compared with that of the individual classifier.

• 한국정보과학회 언어공학연구회:학술대회논문집(한글 및 한국어 정보처리)
• /
• 한국정보과학회언어공학연구회 2023년도 제35회 한글 및 한국어 정보처리 학술대회
• /
• pp.119-124
• /
• 2023

자원이 풍부한 언어를 사용하여 훈련된 모델을 만들고 해당 모델을 사용해 자원이 부족한 언어에 대해 전이 학습하는 방법인 교차언어 전송(Cross-Lingual Transfer)은 다국어 모델을 사용하여 특정한 언어에 맞는 모델을 만들 때 사용되는 일반적이고 효율적인 방법이다. 교차언어 전송의 성능은 서비스하는 언어와 전송 모델을 만들기 위한 훈련 데이터 언어에 따라 성능이 매우 다르므로 어떤 언어를 사용하여 학습할지 결정하는 단계는 효율적인 언어 서비스를 위해 매우 중요하다. 본 연구에서는 교차언어 전송을 위한 원천언어를 찾을 수 있는 특성이 무엇인지 회귀분석을 통해 탐구한다. 또한 교차언어전송에 용이한 원천 학습 언어를 찾는 기존의 방법론들 간의 비교를 통해 더 나은 방법을 도출해내고 한국어의 경우에 일반적으로 더 나은 원천 학습 언어를 찾을 수 있는 방법론을 도출한다.

• Imaging Science in Dentistry
• /
• 제31권2호
• /
• pp.101-107
• /
• 2001

Purpose: To evaluate the precision of measurements of distances and angle in the cross-sectional views of linear tomogram of panorama and to assess the technique for visualizing the mandibular canal. Methods: Ten dry mandibles were radiographically examined with 3 continuous cross-sectional views of linear tomogram of panorama and 4 continuous computed tomograms. The distance between the superior border of canal and alveolar crest and the bucco-lingual width of alveolar bone at the level of the superior border of canal and the angle between the two lines above were measured. Measurements were performed by radiologist and implantologist group and compared with measurements on computed radiograms of the same areas. Results: The measurements differences for the distance of alveolar bone height between in panorama and in CT showed 0.9 mm±0.6 mm by radiologists and 1.3mm±0.8mm by implantologists. There was no statistically significant difference between two groups' measurements. The differences in measurements for the distance of alveolar bone width between in panorama and in CT showed 0.5mm±0.8mm by radiologists and 2.5mm±1.4 mm by implantologists. There was significant difference (P<0.05) between two groups' measurements. The average bucco-lingual inclination of alveolar bone above mandibular canal was average 95.8° in CT. The difference of measurements between two groups was average 1±0.9°. Three cross-sectional views of panorama could show that the mandibular canal crosses antero-lingually and slopes inferiorly from the posterior segment of the mandible. Conclusions: The measurements in the linear tomogram of panorama by radiologists gave the accurate values of the distances and the angle compared with the values in computed tomograms.

• 대한치과교정학회지
• /
• 제48권4호
• /
• pp.236-244
• /
• 2018

Objective: The aim of this study is to compare the adaptation of a straight wire between brackets positioned at the mid-lingual surface and those placed gingivally by using a three-dimensional simulation software. Methods: This cross-sectional study was performed using OrthoAid, an in-house software. The subjects were 36 adolescents with normal Class I occlusion. For each dental cast, two bracket positioning approaches, namely the middle and gingival, were examined. In the middle group, the reference points were placed on the mid-lingual surface of each tooth, while in the gingival group, the reference points were positioned lingually on the anterior teeth. A 4th degree polynomial was adopted, and the in-plane and off-plane root mean squares (RMSs) of the distances between the reference points and the fitted polynomial curve were calculated using the software. Statistical analysis was performed using the paired-samples t-test (${\alpha}=0.05$). Results: The mean in-plane RMS of the polynomial curve to the bracket distance in the gingival group was significantly lower than that in the middle group (p < 0.001). The off-plane RMS was higher in the gingivally positioned brackets in the maxilla than in the middle group (p < 0.001). However, the off-plane RMS in mandible was not statistically significantly different between the two groups (p = 0.274). Conclusions: The results demonstrated that the gingival placement of lingual brackets on the anterior teeth could decrease the distance between a tooth and the straight wire.

• 대한치과보철학회지
• /
• 제30권1호
• /
• pp.25-42
• /
• 1992

The purpose of this study was to analyze the magnitude and distribution of stress using a photoelastic model from a unilateral distal extention removable partial dentures with five kinds of the direct retainers, that is, the bilaterally designed bar clasp of the cross-arch lingual bar and the unilaterally designed bar clasp, circumferential clasp, mini-Dalbo attachment, and telescope retainer. A photoelastic model for mandible was made of the epoxy resin(PL-1) and hardner (PLH-1) with the acrylic resin teeth used and was coated with plastic cement-1 at the lingual surface of the model, and then five kinds of removable partial dentures were set, A unilateral vertical load of about 16Kg was applied on the first molar and the stress pattern of the photoelastic model under each condition was analyzed by the reflective circular polariscope. The following results were obtained: 1. The conventional removable partial denture with the bilaterally cross arch lingual bar produced the most favorable stress distribution on the residual ridge and supporting structure of abutment teeth than the unilaterally designed removable partial dentures. 2. The unilaterally designed removable partial denture with the bar clasp produced the stress distribution on the residual ridge, except sligtly higher stress concentration on the supporting structure of the abutment teeth, similar to the conventional removable partial denture with the bilaterally designed cross arch lingual bar. 3. On the unilaterally designed removable partial dentures, the bar clasp produced greater stress distribution on the residual ridge and supporting structure of the abutment teeth than the circumferential clasp. 4. On the unilaterally designed removable partial dentures, the mimi-Dalbo attachment produced relatively higher stress concentration on the residual ridge, but produced lesser stress concentration on the supporting structure of the abutment teeth than the other direct retainers. 5. On the unilaterally designed removable partial dentures, the telescope retainer produced uniform stress distribution on the residual ridge, but produced higher stress concentration at the root apex of the terminal abutment tooth than the other direct retainers. 6. On the unilaterally designed removable partial dentures the circumferential clasp and telescope retainer produced slightly higher stress concentration on the residual ridge and supporting structure of the abutment teeth than the bar clasp and mini- Dalbo attachment.

• 대한치과보철학회지
• /
• 제34권4호
• /
• pp.746-754
• /
• 1996

개교증, 절연부 파절, 변색등 심미적 요인으로 라미네이트를 설측연장시 연장길이에 따른 파절강도를 알아보기위해 발거된 60개의 상악 전치의 순면에 3개의 seating form을 형성하고sheet wax를 사용하여 가로 5mm $^*$세로 5mm $^*$두께 0.9mm의 block을 형성하고, 각 군에 따라 I군 : 0.5mm, II군 : 1mm, III군 : 2mm, IV군 : 0mm로 설측 연장 길이를 달리하여 block을 형성 이를 인상채득하여 내화모형을 제작하였다. 여기에 라미네이트용 도재를 2회에 걸쳐 축성 통법대로 소성 제작하였다. block에 맞추어 contouring하고, 내화매몰재를 제거 sandblast처리를 시행하였다. 그후 레진시멘트로 합착하였다. 절치간 각도를 127도로 특별히 제작된 holding device를 이용하여 인스트론 만능시험기에 시편을 거상하고, 분당 0.5mm cross head speed로 파절강도를 측정하여 다음과 같은 결과를 얻었다. 1. 각 군의 평균파절강도는 대조군인 설면으로 연장치 않은 IV군에서 86.95Mpa, 0.5mm 연장한 I군에서 44.98Mpa, 1.0mm연장한 II군에서 27.47Mpa, 2.0mm연장한 III군에서 19.61Mpa의 순으로 나타났다. 2. 모든 실험군 사이에는 통계적으로 유의성 있는 차이를 나타냈고(p<0.01), 이러한 실험 결과로 미루어 보아 라미네이트 디자인시, 설면으로 연장치 않거나, 절연부의 피개가 불가피한 경우에는 설면으로 0.5mm 연장하는 것이 파절 강도면에서 유리하다고 사료된다.

• Imaging Science in Dentistry
• /
• 제37권4호
• /
• pp.217-220
• /
• 2007

Purpose: The purpose of this study was to compare the location of the mandibular canal in Class III malocclusion to its location in normal occlusion for adults. Materials and Methods: For this study 32 skeletal Class III patients and 26 normal patients were observed. Four measurements were taken on cross sectional tomography between the first and second molars: the distance from the mandibular canal to the inner surface of both the buccal and lingual cortices, the distance from the mandibular canal to the inferior border of the mandible, and the buccolingual width of the mandible. The buccolingual location of the canals was classified as lingual, central, or buccal. Each measurement was analyzed with an independent t test to compare Class III malocclusion to normal occlusion. Results: Compared to the control group, the prognathic group had a shorter distance from the canal to the inner surface of the lingual cortex and to the base of the mandible. A higher percentage of the canals were located lingually in the prognathic group. Conclusion: This study showed that the mandibular canal was located more lingually and inferiorly in prognathic patients than in patients with normal occlusion. These results could help surgeons to reduce injuries to the inferior alveolar nerve.