• The Journal of Korean Institute of Communications and Information Sciences
• /
• v.26 no.7B
• /
• pp.957-967
• /
• 2001

본 논문에서는 모델 기반 지각적 그룹핑을 이용한 3차원 다면체의 인식 및 형상 복원에 관한 새로운 기법을 제안한다. 2차원 입력 영상과 여기에서 추출된 특징들의 3차원 특징을 거리 측정기를 이용하여 추출하여 인식 및 복원의 기본 특징으로 이용한다. 이 때, 모델의 3차원 기하학적 정보는 결정 트리 분류기에 의하여 학습되며 지각적 그룹핑은 이와 같은 모델 기반으로 이루어진다. 또한, 1차 그룹핑의 결과로 얻어진 3차원 직선 특징간의 관계는 Gestalt 그래프로 표현되며 이것의 부그래프 분할을 통하여 인식을 위한 후보 그룹이 생성된다. 마지막으로 각각의 후보 그룹은 3차원 모델과 정렬되어 가장 잘 부합되는 그룹을 인식 결과로 생성하게 된다. 그리고 정렬의 결과로서 2차원 텍스춰를 추출하여 3차원 모델에 매핑함으로써 실제적인 3차원 형상을 복원할 수 있다. 제안하는 알고리듬의 성능을 평가하기 위하여 불록 영상과 지형 모델 보드 영상에 대하여 실험을 수행하였다. 실험 결과, 모델 기반의 그룹핑 기법은 결과 그룹의 수를 상당히 감소시켰으며 또한 잡음과 가리워짐에 강건한 인식과 복원 결과가 얻어졌다.

• Annual Conference on Human and Language Technology
• /
• 2021.10a
• /
• pp.427-432
• /
• 2021

현재 한국어 의존 구문 분석의 표준은 어절 단위로 구문 분석을 수행하는 것이다. 그러나 의존 구문 분석의 분석 단위(어절, 형태소)에 대해서는 현재까지 심도 있는 비교 연구가 진행된 바 없다. 본 연구에서는 의존 구문 분석의 분석 단위가 자연어 처리 분야의 성능에 유의미한 영향을 끼침을 실험적으로 규명한다. STEP 2000과 모두의 말뭉치를 기반으로 구축한 형태소 단위 의존 구문 분석 말뭉치를 사용하여, 의존 구문 분석기 모델 및 의존 트리를 입력으로 활용하는 문장 의미 유사도 분석(STS) 및 관계 추출(RE) 모델을 학습하였다. 그 결과, KMDP가 기존 어절 단위 구문 분석과 비교하여 의존 구문 분석기의 성능과 응용 분야(STS, RE)의 성능이 모두 유의미하게 향상됨을 확인하였다. 이로써 형태소 단위 의존 구문 분석이 한국어 문법을 표현하는 능력이 우수하며, 문법과 의미를 연결하는 인터페이스로써 높은 활용 가치가 있음을 입증한다.

• Korean Journal of Cognitive Science
• /
• v.4 no.2
• /
• pp.85-114
• /
• 1994

One of the major research topics in the area of the development of intelligent tutoring system(ITS)is the control of instructional mechanism consisting of lesson plans,curriculum plans,and discourse plans.This paper describes a method of building the lesson plans among these three instructional plans based on the case-based planning.It is more efficient to retrieve the lesson plan from the plan memoru than to generate it whenever an instructional goal is selected.The retrieved lesson plan may be modified to build more adaptive plan for the current goal.We have developed a lesson plan generator that has such capabilities as a component of an ITS for teching indefinite intergration.We also have devised a description language to represeint the generalized form for the given arithmetic expression as an instructional goal and a curriculum tree to represent the lesson units required to master the subject matter.The result of this research could be used either by a developer of the lesson plan generator in the other area of ITS or by human teacher as a curriculum in the actual class.

• Journal of KIISE:Computing Practices and Letters
• /
• v.15 no.5
• /
• pp.345-349
• /
• 2009

Social bookmarking systems are a typical web 2.0 service based on folksonomy, providing the platform for storing and sharing bookmarking information. Spammers in social bookmarking systems denote the users who abuse the system for their own interests in an improper way. They can make the entire resources in social bookmarking systems useless by posting lots of wrong information. Hence, it is important to detect spammers as early as possible and protect social bookmarking systems from their attack. In this paper, we applied a diverse set of machine learning approaches, i.e., decision tables, decision trees (ID3), $na{\ddot{i}}ve$ Bayes classifiers, TAN (tree-augment $na{\ddot{i}}ve$ Bayes) classifiers, and artificial neural networks to this task. In our experiments, $na{\ddot{i}}ve$ Bayes classifiers performed significantly better than other methods with respect to the AUC (area under the ROC curve) score as veil as the model building time. Plausible explanations for this result are as follows. First, $na{\ddot{i}}ve$> Bayes classifiers art known to usually perform better than decision trees in terms of the AUC score. Second, the spammer detection problem in our experiments is likely to be linearly separable.

• Journal of the Institute of Electronics and Information Engineers
• /
• v.54 no.6
• /
• pp.100-105
• /
• 2017

This work presents an ECG biometric recognition system for the purpose of biometric authentication. ECG biometric approaches are divided into two major categories, fiducial-based and non-fiducial-based methods. This paper proposes a new non-fiducial framework using discrete cosine transform and a Random Forest classifier. When using DCT, most of the signal information tends to be concentrated in a few low-frequency components. In order to apply feature vector of Random Forest, DCT feature vectors of ECG heartbeats are constructed by using the first 40 DCT coefficients. RF is based on the computation of a large number of decision trees. It is relatively fast, robust and inherently suitable for multi-class problems. Furthermore, it trade-off threshold between admission and rejection of ID inside RF classifier. As a result, proposed method offers 99.9% recognition rates when tested on MIT-BIH NSRDB.

• The Journal of the Acoustical Society of Korea
• /
• v.23 no.2
• /
• pp.173-179
• /
• 2004

Due to the insufficiency of training data in large vocabulary continuous speech recognition, similar context dependent phones can be clustered by decision trees to share the data. When the decision trees are built and used to predict unseen triphones, a phonetic question set is required. The phonetic question set, which contains categories of the phones with similar co-articulation effects, is usually generated by phonetic or linguistic experts. This knowledge-based approach for generating phonetic question set, however, may reduce the homogeneity of the clusters. Moreover, the experts must adjust the question sets whenever the language or the PLU (phone-like unit) of a recognition system is changed. Therefore, we propose a data-driven method to automatically generate phonetic question set. Since the proposed method generates the phone categories using speech data distribution, it is not dependent on the language or the PLU, and may enhance the homogeneity of the clusters. In large vocabulary speech recognition experiments, the proposed algorithm has been found to reduce the error rate by 14.3%.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.15 no.3
• /
• pp.289-293
• /
• 2005

Pattern classification extracts various types of pattern information expressing objects in the real world and decides their class. The top priority of pattern classification technologies is to improve the performance of classification and, for this, many researches have tried various approaches for the last 40 years. Classification methods used in pattern classification include base classifier based on the probabilistic inference of patterns, decision tree, method based on distance function, neural network and clustering but they are not efficient in analyzing a large amount of multi-dimensional data. Thus, there are active researches on multiple classifier systems, which improve the performance of classification by combining problems using a number of mutually compensatory classifiers. The present study identifies problems in previous researches on multiple SVM classifiers, and proposes BORSE, a model that, based on 1:M policy in order to expand SVM to a multiple class classifier, regards each SVM output as a signal with non-linear pattern, trains the neural network for the pattern and combine the final results of classification performance.

• The KIPS Transactions:PartB
• /
• v.15B no.4
• /
• pp.375-382
• /
• 2008

A protein's subcellular localization is considered an essential part of the description of its associated biomolecular phenomena. As the volume of biomolecular reports has increased, there has been a great deal of research on text mining to detect protein subcellular localization information in documents. It has been argued that linguistic information, especially syntactic information, is useful for identifying the subcellular localizations of proteins of interest. However, previous systems for detecting protein subcellular localization information used only shallow syntactic parsers, and showed poor performance. Thus, there remains a need to use a full syntactic parser and to apply deep linguistic knowledge to the analysis of text for protein subcellular localization information. In addition, we have attempted to use semantic information from the WordNet thesaurus. To improve performance in detecting protein subcellular localization information, this paper proposes a three-step method based on a full syntactic dependency parser and WordNet thesaurus. In the first step, we constructed syntactic dependency paths from each protein to its location candidate, and then converted the syntactic dependency paths into dependency trees. In the second step, we retrieved root information of the syntactic dependency trees. In the final step, we extracted syn-semantic patterns of protein subtrees and location subtrees. From the root and subtree nodes, we extracted syntactic category and syntactic direction as syntactic information, and synset offset of the WordNet thesaurus as semantic information. According to the root information and syn-semantic patterns of subtrees from the training data, we extracted (protein, localization) pairs from the test sentences. Even with no biomolecular knowledge, our method showed reasonable performance in experimental results using Medline abstract data. Our proposed method gave an F-measure of 74.53% for training data and 58.90% for test data, significantly outperforming previous methods, by 12-25%.

• The Journal of the Korea institute of electronic communication sciences
• /
• v.18 no.5
• /
• pp.999-1008
• /
• 2023

Early prediction of chronic diseases such as diabetes is an important issue, and improving the accuracy of diabetes prediction is especially important. Various machine learning and deep learning-based methodologies are being introduced for diabetes prediction, but these technologies require large amounts of data for better performance than other methodologies, and the learning cost is high due to complex data models. In this study, we aim to verify the claim that DNN using the pima dataset and k-fold cross-validation reduces the efficiency of diabetes diagnosis models. Machine learning classification methods such as decision trees, SVM, random forests, logistic regression, KNN, and various ensemble techniques were used to determine which algorithm produces the best prediction results. After training and testing all classification models, the proposed system provided the best results on XGBoost classifier with ADASYN method, with accuracy of 81%, F1 coefficient of 0.81, and AUC of 0.84. Additionally, a domain adaptation method was implemented to demonstrate the versatility of the proposed system. An explainable AI approach using the LIME and SHAP frameworks was implemented to understand how the model predicts the final outcome.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.13 no.4
• /
• pp.805-812
• /
• 2009

In many intrusion detection applications, n-gram approach has been widely applied. However, n-gram approach has shown a few problems including unscalability and double counting of features. To address those problems, we applied n-gram augmented Naive Bayes with k-truncated suffix tree (k-TST) storage mechanism directly to classify intrusive sequences and compared performance with those of Naive Bayes and Support Vector Machines (SVM) with n-gram features by the experiments on host-based intrusion detection benchmark data sets. Experimental results on the University of New Mexico (UNM) benchmark data sets show that the n-gram augmented method, which solves the problem of independence violation that happens when n-gram features are directly applied to Naive Bayes (i.e. Naive Bayes with n-gram features), yields intrusion detectors with higher accuracy than those from Naive Bayes with n-gram features and shows comparable accuracy to those from SVM with n-gram features. For the scalable and efficient counting of n-gram features, we use k-truncated suffix tree mechanism for storing n-gram features. With the k-truncated suffix tree storage mechanism, we tested the performance of the classifiers up to 20-gram, which illustrates the scalability and accuracy of n-gram augmented Naive Bayes with k-truncated suffix tree storage mechanism.