• ETRI Journal
• /
• v.23 no.4
• /
• pp.168-176
• /
• 2001

This paper discusses two important issues of corpus-based synthesis: synthesis unit generation based on phrase break strength information and pruning redundant synthesis unit instances. First, the new sentence set for recording was designed to make an efficient synthesis database, reflecting the characteristics of the Korean language. To obtain prosodic context sensitive units, we graded major prosodic phrases into 5 distinctive levels according to pause length and then discriminated intra-word triphones using the levels. Using the synthesis unit with phrase break strength information, synthetic speech was generated and evaluated subjectively. Second, a new pruning method based on weighted vector quantization (WVQ) was proposed to eliminate redundant synthesis unit instances from the synthesis database. WVQ takes the relative importance of each instance into account when clustering similar instances using vector quantization (VQ) technique. The proposed method was compared with two conventional pruning methods through objective and subjective evaluations of synthetic speech quality: one to simply limit the maximum number of instances, and the other based on normal VQ-based clustering. For the same reduction rate of instance number, the proposed method showed the best performance. The synthetic speech with reduction rate 45% had almost no perceptible degradation as compared to the synthetic speech without instance reduction.

• Journal of the Korean Society for information Management
• /
• v.23 no.1 s.59
• /
• pp.279-299
• /
• 2006

This study utilizes various approaches of machine learning in the process of automatically assigning descriptors to journal articles. The effectiveness of feature selection and the size of training set were examined, after selecting core journals in the field of information science and organizing test collection from the articles of the past 11 years. Regarding feature selection, after reducing the feature set using $x^2$ statistics(CHI) and criteria that prefer high-frequency features(COS, GSS, JAC), the trained Support Vector Machines(SVM) performed the best. With respect to the size of the training set, it significantly influenced the performance of Support Vector Machines(SVM) and Voted Perceptron(VTP). However, it had little effect on Naive Bayes(NB).

• Journal of the Korea Society of Computer and Information
• /
• v.23 no.7
• /
• pp.19-26
• /
• 2018

This paper presents a structuring process of unstructured social network service (SNS) messages on rail services. We crawl messages about rail services posted on SNS and extract keywords indicating date and time, rail operating company, station name, direction, and rail service types from each message. Among them, the rail service types are classified by machine learning according to predefined rail service types, and the rest are extracted by regular expressions. Words are converted into vector representations using Word2Vec and a conventional Convolutional Neural Network (CNN) is used for training and classification. For performance measurement, our experimental results show a comparison with a TF-IDF and Support Vector Machine (SVM) approach. This structured information in the database and can be easily used for services for railway users.

• Annual Conference on Human and Language Technology
• /
• 2001.10d
• /
• pp.209-212
• /
• 2001

본 논문에서는 text-to-speech 시스템에서 사용할 억양 모델을 위해 벡터 양자화(vector quantization) 방식을 이용한다. 어절 경계강도(break index)는 세단계로 분류하였고, CART(Classification And Regression Tree)를 사용하여 어절 경계강도의 예측 규칙을 생성하였다. 예측된 어절 경계강도를 바탕으로 운율구를 예측하였으며 운율구는 다섯 개의 억양 패턴으로 분류하였다. 하나의 운율구는 정점(peak)의 시간축, 주파수축 값과 이를 기준으로 한 앞, 뒤 기울기를 추출하여 네 개의 파라미터로 단순화하였다. 운율구에 대해서 먼저 운율구가 문장의 끝일 경우와 아닐 경우로 분류하고, 억양 패턴 다섯 개로 분류하여. 모두 10개의 운율구 set으로 나누었다. 그리고 네 개의 파라미터를 가지고 있는 운율구의 억양 패턴을 벡터 양자화 방식을 이용하여 분류(clusteing)하였다 운율의 변화가 두드러지는 조사와 어미는 12 point의 기본주파수 값을 추출하고 벡터 양자화하였다. 운율구와 조사 어미의 codebook index는 문장에 대한 특징 변수 값을 추출하고 CART를 사용하여 예측하였다. 합성할 때에는 입력 tort에 대해서 운율구의 억양 파라미터를 추정한 다음, 조사와 어미의 12 point 기본주파수 값을 추정하여 전체 억양 곡선을 생성하였고 본 연구실에서 제작한 음성합성기를 통해 합성하였다.

• Journal of the Korean Society of Safety
• /
• v.32 no.5
• /
• pp.179-185
• /
• 2017

Accident analysis is an essential process to make basic data for accident prevention. Most researches depend on survey data and accident statistics to analyze accidents, but these kinds of data are not sufficient for systematic and detailed analysis. We, in this paper, propose an accident classification model that extracts task type, original cause materials, accident type, and the number of deaths from accident reports. The classification model is a support vector machine (SVM) with word occurrence features, and these features are selected based on mutual information. Experiment shows that the proposed model can extract task type, original cause materials, accident type, and the number of deaths with almost 100% accuracy. We also develop an accident ontology to express the information extracted by the classification model. Finally, we illustrate how the proposed classification model and ontology effectively works for the accident analysis. The classification model and ontology are expected to effectively analyze various accidents.

• Journal of KIISE:Computing Practices and Letters
• /
• v.14 no.3
• /
• pp.336-340
• /
• 2008

Sentiment classification is a recent subdiscipline of text classification, which is concerned not with the topic but with opinion. In this paper, we present a Korean sentence and document classification system using effective sentiment features. Korean sentiment classification starts from constructing effective sentiment feature sets for positive and negative. The synonym information of a English word thesaurus is used to extract effective sentiment features and then the extracted English sentiment features are translated in Korean features by English-Korean dictionary. A sentence or a document is represented by using the extracted sentiment features and is classified and evaluated by SVM(Support Vector Machine).

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.8 no.7
• /
• pp.1546-1552
• /
• 2004

Electronic Ink is a stored data in the form of the handwritten text or the script without converting it into ASCII by handwritten recognition on the pen-based computers and Personal Digital Assistants(Pocket PC) for supporting natural and convenient data input. One of the most important issues is to search the electronic ink in order to use it. We proposed and implemented a script matching algorithm for the electronic ink. Proposed matching algorithm separated the input stroke into a set of primitive stroke using the curvature of the stroke curve. After determining the type of separated strokes, it produced a stroke feature vector. And then it calculated the distance between the stroke feature vector of input strokes and one of strokes in the database using the dynamic programming technique.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.21 no.3
• /
• pp.45-56
• /
• 2011

Hyper Text Transfer Protocol(HTTP) is widely used in nearly every network when people access web pages, therefore HTTP traffic is usually allowed by local security policies to pass though firewalls and other gateway security devices without examination. However this characteristic can be used by malicious people. With the help of HTTP tunnel applications, malicious people can transmit data within HTTP in order to circumvent local security policies. Thus it is quite important to distinguish between regular HTTP traffic and tunneled HTTP traffic. Our work of HTTP tunnel detection is based on Support Vector Machines. The experimental results show the high accuracy of HTTP tunnel detection. Moreover, being trained once, our work of HTTP tunnel detection can be applied to other places without training any more.

• Journal of Internet Computing and Services
• /
• v.18 no.6
• /
• pp.35-46
• /
• 2017

Recently, attackers using malicious code in cyber security have been increased by attaching malicious code to a mail and inducing the user to execute it. Especially, it is dangerous because it is easy to execute by attaching a document type file. The author analysis is a research area that is being studied in NLP (Neutral Language Process) and text mining, and it studies methods of analyzing authors by analyzing text sentences, texts, and documents in a specific language. In case of attack mail, it is created by the attacker. Therefore, by analyzing the contents of the mail and the attached document file and identifying the corresponding author, it is possible to discover more distinctive features from the normal mail and improve the detection accuracy. In this pager, we proposed IADA2(Intelligent Attack mail Detection based on Authorship Analysis) model for attack mail detection. The feature vector that can classify and detect attack mail from the features used in the existing machine learning based spam detection model and the features used in the author analysis of the document and the IADA2 detection model. We have improved the detection models of attack mails by simply detecting term features and extracted features that reflect the sequence characteristics of words by applying n-grams. Result of experiment show that the proposed method improves performance according to feature combinations, feature selection techniques, and appropriate models.

• Korean Journal of Biological Psychiatry
• /
• v.27 no.1
• /
• pp.18-26
• /
• 2020

Objectives The aim was to find effective vectorization and classification models to predict a psychiatric diagnosis from text-based medical records. Methods Electronic medical records (n = 494) of present illness were collected retrospectively in inpatient admission notes with three diagnoses of major depressive disorder, type 1 bipolar disorder, and schizophrenia. Data were split into 400 training data and 94 independent validation data. Data were vectorized by two different models such as term frequency-inverse document frequency (TF-IDF) and Doc2vec. Machine learning models for classification including stochastic gradient descent, logistic regression, support vector classification, and deep learning (DL) were applied to predict three psychiatric diagnoses. Five-fold cross-validation was used to find an effective model. Metrics such as accuracy, precision, recall, and F1-score were measured for comparison between the models. Results Five-fold cross-validation in training data showed DL model with Doc2vec was the most effective model to predict the diagnosis (accuracy = 0.87, F1-score = 0.87). However, these metrics have been reduced in independent test data set with final working DL models (accuracy = 0.79, F1-score = 0.79), while the model of logistic regression and support vector machine with Doc2vec showed slightly better performance (accuracy = 0.80, F1-score = 0.80) than the DL models with Doc2vec and others with TF-IDF. Conclusions The current results suggest that the vectorization may have more impact on the performance of classification than the machine learning model. However, data set had a number of limitations including small sample size, imbalance among the category, and its generalizability. With this regard, the need for research with multi-sites and large samples is suggested to improve the machine learning models.