Earticle

This study aimed to measure the rate of speaker appearances by examining video footage of individualized interviews from the television program, Tales of Modern Korean History: TV Autobiographies. The appearance rate was calculated based on the possibility of observing the speaker’s facial expressions. Speakers appeared in approximately 80 per cent of all conversations, however, the ratio varied according to the role of the speaker and type of speech. The conversations were classified into three speech categories: expressing agreement, questioning, and expressing neutral opinions. The rate of expression of agreement was the lowest compared to the other categories while the rate for questioning was similar to that of expression of neutral opinions. Moreover, the rate of speaker appearance for all categories was higher for the interviewee than the interviewer. The proportion of a person’s appearance within the video footage also differs for the interviewer and interviewee, with a full body shot of the speaker often used at the beginning of the questioning process. Additional studies need to be conducted on other television programs, to confirm the generalization of this trend specific to Korean interview-based television programs.

본 논문에서는 Sentence-BERT 바이-인코더에 기반한 단문 분류 방법을 제안하고, 이를 초등 인공지능 프로그래 밍 교육 환경에서 텍스트 학습 및 추론 시 효과적으로 적용할 수 있음을 실험적으로 보인다. 본 단문 분류 방법은 신경망을 실제로 처음부터 학습시키거나 미세 조정을 하지 않기 때문에, ‘가상-학습’ 및 ‘가상-추론’ 방법이라고 명 명하였다. 가상-학습 단계에서는 학습할 단문들을 경량 바이-인코더를 이용하여 임베딩하고, 이를 효과적으로 검색 할 수 있도록 인덱스를 생성한다. 가상-추론 단계에서는 추론할 단문들을 동일한 바이-인코더를 이용하여 임베딩하 고, FAISS, HNSW 등과 같은 근사 인접 이웃 알고리즘을 활용하여 빠르게 유사한 임베딩 벡터를 찾아 레이블 예 측을 수행한다. 이 방식을 초등 인공지능 교육에서 많이 활용되는 LSTM과 비교한 결과, 클래스 별 학습 샘플의 수 가 많지 않은 상황에서 본 방식은 LSTM보다 정확도가 크게 앞서는 것으로 나타났다. 그리고 학생들이 저성능의 컴퓨터를 활용하는 환경에서 충분히 활용할 수 있는 수준으로 학습 시간과 추론 시간이 소요됨을 확인하였다.

In this paper, we present a short sentence classification approach based on a Sentence-BERT bi-encoder and experimentally demonstrate its effectiveness for training text data in elementary AI programming educational environments. Because this approach does not train from scratch or fine-tune a neural network, we have termed it the "pseudo-training" and "pseudo-inference" method. In the pseudo-training phase, the short sentences to be trained are embedded using a lightweight bi-encoder, and an index is generated for efficient retrieval. In the pseudo-inference phase, the short sentences to be inferred are embedded using the same bi-encoder, and approximate nearest neighbor algorithms such as FAISS and HNSW are utilized to quickly find similar embedding vectors, followed by label prediction. Compared to LSTM, which is widely used in elementary AI education, the proposed method showed higher accuracy in situations where the number of training samples per class is limited. Additionally, we confirmed that the training and inference time for our approach is sufficient for practical use in programming environments where students use low-performance computers.

본 연구는 수출보험 종목 중 단기수출보험에서 사고발생가능성과 관련하여 분류하는 수출자 등급, 수입자 등급, 수입국 등급 등의 기준이 보험인수건과 사고건을 적절하게 구분하고 있는지를 로지스틱회귀모형을 사용하여 분석하고자 하였다. 분석결과 L/C 건을 포함한 모형과 L/C 건을 제외한 모형의 결과는 비슷하게 나타났으나 전반 적으로 L/C건을 제외한 모형에서 L/C 건을 포함한 모형보다 사고발생의 확률이 낮은 것으로 분석 되었다. 변수별로는 수출자 등급, 수입자 등급, 결제조건, 보험요율 등은 사고건과 무사고건을 유의하게 구분하는 하나의 기준으로 작용했으나 기업규모와 보험종목은 유효한 기준으로 작용하 지 못하였다.

The factors such as exporter and importer ratings , import country grade etc. are designed to properly classify the possibility of loss in the short-term export credit insurance. This paper analyzes whether these factors are well- designed or not in the respect of risk classification by using a logistic regression model. The research results show that exporter ratings, importer ratings, settlement terms, insurance premium rate are valid criteria to classify the accident occurrence but the size of firms and short- term products are not. This results are similar between L / C including model and L / C excluding model but the possibility of accident usually shows lower in L / C excluding the model.

본 논문에서는 이미지 학습 및 분류에 가장 많이 활용되고 있는 CNN모델중의 하나인 VGG19알고리즘 을 이용하여 전기화재시 발생하는 용융 흔적을 분류를 방법을 제안하였다. 전기화재의 직접적인 유무를 감지하기 위한 가장 기본적인 근거는 전선의 녹는 정도와 형태에 따라 전기화재 현장에서 발생하는 전선의 용융흔적이다. 제 안된 방법은 VGG19 알고리즘을 활용하여 용융 흔적의 유무를 검출할 수 있도록 변형하여 사용하였으며, 학습에 필요한 데이터는 연구실에서 실제제적하여 사용하였다. 실험결과 제안된 알고리즘의 최종 검증 정확도는 96.31% 이고, 손실율은 0.1169로 확인되었다. 이와 같이 높은 정확도를 확보한 결과를 통하여 용융흔적 검출 알고리즘이 전기화재 유무 검증에 활용될 수 있는 가능성을 보였다.

In this paper, the VGG19 algorithm was used by applying the transfer learning for the classification of molten traces of electric fire arc-beads data which is one of the most used models in convolutional neural network(CNN) computer vision tasks. The most essential basis for detecting direct indications of electric fires is the melting traces of wires that occur at the site of an electric fire, depending on the severity and shape of the melting. The proposed VGG19 method was altered and used such that it could detect molten traces, and the molten trace data of the wires required for learning were created in the lab. The final validation accuracy result was 96.31% with validation loss of 0.1169. Through the result of securing such high accuracy, the possibility of using the melting trace detection algorithm to verify the presence or absence of an electric fire was shown.

자연어 처리를 이용한 문서 분류 분야에서도 전통적인 방법에서 벗어나 단어 임베딩을 활용한 합성곱 신경망과 순환 신경망 등 심층 신경망을 이용한 다양한 연구가 진행되고 있다. 본 논문에서는 Word2Vec과 두 개의 계층으로 구성 된 양방향 장단기 기억 네트워크를 이용한 특허 문서의 IPC(International Patents Classification) 자동 분류 모델을 제안한다. IPC는 세계지식재산권기구에서 제정한 국제적으로 통일된 특허 분류 기준이며, 각 국가의 공인된 기관에서 수작업으로 분류하고 있다. IPC 자동 분류를 위하여 입력 시퀀스에 Word2Vec을 이용한 단어 임베딩가 중치를 사용한다. 그리고 가중치가 부여된 시퀀스를 두 개의 계층을 갖는 깊은 구조의 양방향 장단기 기억 네트워크 신경망에 입력하여 IPC를 분류한다. 실험 결과 특허 문서의 분류 정확도가 합성곱 신경망 보다는 약 7% 향상되었 으며, 순환 신경망을 단일로 이용하는 것 보다는 약 5% 향상된 것을 확인할 수 있었다. 또한 전통적인 방법인 나이 브 베이시안, 로지스틱 분류 및 서포트 벡터 머신보다는 5~12% 이상 우수한 성능을 나타내었다.

There are various studies using Deep Neural Network such as CNN(Convolutional Neural Network) and RNN(Recurrent Neural Network) that utilize word embedding in document classification using natural language processing out of traditional methods. In this paper, we propose the IPC(International Patents Classification) automatic classification model of patent documents using two layers BLSTM (Bidirectional Long Short Term memory) network. The IPC is an internationally uniform standard for patent classification established by the World Intellectual Property Organization and is categorized by hand in authorized agencies in each country. For the IPC automatic classification, we use word embedding weight with Word2Vec in the input sequences. And they are classified by entering a weighted sequences into a deep neural network with two layers BLSTM. The experimental results showed that the accuracy of classification is improved by about 7% than that of CNN, and about 5% than that of single layer LSTM that is a field of RNN. Also it showed more than 5~12% higher performance than traditional methods such as Naive Bayes, Logistic and Support Vector Machine classification.

Short text is a popular text form, which is widely used in real-time network news, short commentary, micro-blog and many other fields. With the development of the application such as QQ, mobile phone text messages and movie websites, the size of data is also becoming larger and larger. Most data is useless for us while other data is significant for us. Therefore, it is necessary for us to extract the useful short text from the big data. However, there are many problems with the short text classification, such as fewer features, irregularity and so on. To solve these problems, we should pretreat the short text set first, and then choose the significant features. This paper use semi-supervised learning method and SVM classifier to improve the traditional methods and it can classify a large number of short texts to mining the useful massage from the short text. The experimental results in this paper also show a good promotion.

The noise processing is the key of improving recognition rate for the noisy utterance. While for the short utterance, its corpus is less and small amount of speech data is available for testing and training, so making full use of its limit corpus is the key of improving recognition rate of the short utterance. For the noisy short utterance, the noise processing and making full use of the limit corpus are vital. We proposed noise separation algorithm based on constrained Non-negative matrix factorization (CNMF) to make the noise processing. As making full use of the limit corpus, we proposed the improved SNR discrimination algorithm (ISNRDA) and the differences detection and discrimination algorithm (DDADA), we use the two classification algorithm to estimate the quality of the speech frame, and classify the speech frame. Besides, we combine the above classification result with the GMM-UBM three-stage classification model proposed in this paper, so that we can make full use of the limit corpus of the noisy short utterance. Experiments show that the above algorithms can improve speaker recognition performance of noisy short utterance.

The noisy short utterance is polluted by noise and corpus is less, so the recognition rate significantly decreased. For improving recognition rate, we proposed the dual information quality discrimination algorithm to classify the speech frames: one is differences detection and discrimination algorithm (DDADA), another is the improved SNR discrimination algorithm (ISNRDA). Based on the above two algorithms, the speech frames are classified to three classes: high quality, medium quality and low quality. We proposed GMM-UBM three-stage classification model, and we combine the dual information quality discrimination algorithm with GMM-UBM three-stage classification model. Experiments show that, the dual discrimination quality algorithms can be more precise to classify speech frame, and combining it with GMM-UBM three-stage classification model can make full use of limited corpus of short utterance and can improve the speaker recognition rate of the noisy short utterance.

데이터 마이닝의 문서분류 기술에서 발전된 오피니언 마이닝은 이제 국외뿐만 아니라 국내 산업에서 중요한 관심분야로 자리잡아가고 있다. 오피니언 마이닝의 핵심은 문서에서 감정 단어를 추출하여 긍정/부정 여부를 얼마나 정확하게 판별하느냐를 평가하는 것이다. 국내에서도 이에 관련된 많은 연구가 이루어 졌으나 아직 실용적으로 적용 할 만큼의 분류 정확도를 보이지 않고 있다. 한국어의 경우 비문법적 표현, 감정단어의 다양성 등으로 인해 문서의 극성을 판별하기가 쉽지 않기 때문이다. 본 논문에서는 문법적 요소를 최대한 배제하고 단어패턴의 빈도만을 고려한 새로운 오피니언 문서 분류기법을 제안한다. 제안된 방법에서는 문서를 단어들의 리스트로 추상화한 후, 패턴들의 빈 도를 이용하여 기계학습 알고리즘을 적용한다. 이후에 적절한 스코어 함수를 적용하여 문서의 극성을 판별한다. 또한 제안된 기법의 정확도를 평가하기 위해서 실험결과를 제시한다.

An opinion mining technique which was developed from document classification in area of data mining now becomes a common interest in domestic as well as international industries. The core of opinion mining is to decide precisely whether an opinion document is a positive or negative one. Although many related approaches have been previously proposed, a classification accuracy was not satisfiable enough to applying them in practical applications. A opinion documents written in Korean are not easy to determine a polarity automatically because they often include various and ungrammatical words in expressing subjective opinions. Proposed in this paper is a new approach of classification of opinion documents, which considers only a frequency of word patterns and excludes the grammatical factors as much as possible. In proposed method, we express a document into a bag of words and then apply a learning algorithm using a frequency of word patterns, and finally decide the polarity of the document using a score function. Additionally, we also present the experiment results for evaluating the accuracy of the proposed method.

The research related to fault diagnosis in permanent magnet synchronous motors (PMSMs) has attracted considerable attention in recent years because various faults such as permanent magnet demagnetization and short-circuited turns can occur and result in unexpected failure of motor related system. Several conventional current and back electromotive force (BEMF) analysis techniques were proposed to detect certain faults in PMSMs; however, they generally deal with a single fault only. On the contrary, cases of multiple faults are common in PMSMs. We propose a fault diagnosis method for PMSMs with single and multiple combined faults. Our method uses three phase BEMF voltages based on the fast Fourier transform (FFT), support vector machine(SVM), and visualization tools for identifying fault types and severities in PMSMs. Principal component analysis (PCA) and t-distributed stochastic neighbor embedding (t-SNE) are used to visualize the high-dimensional data into two-dimensional space. Experimental results show good visualization performance and high classification accuracy to identify fault types and severities for single and multiple faults in PMSMs.

Objectives: Both the valence and arousal components of affect are important considerations when managing mental healthcare because they are associated with affective and physiological responses. Research on arousal and valence analysis, which uses images, texts, and physiological signals that employ deep learning, is actively underway; research investigating how to improve the recognition rate is needed. The goal of this research was to design a deep learning framework and model to classify arousal and valence, indicating positive and negative degrees of emotion as high or low. Methods: The proposed arousal and valence classification model to analyze the affective state was tested using data from 40 channels provided by a dataset for emotion analysis using electrocardiography (EEG), physiological, and video signals (the DEAP dataset). Experiments were based on 10 selected featured central and peripheral nervous system data points, using long short-term memory (LSTM) as a deep learning method. Results: The arousal and valence were classified and visualized on a two-dimensional coordinate plane. Profiles were designed depending on the number of hidden layers, nodes, and hyperparameters according to the error rate. The experimental results show an arousal and valence classification model accuracy of 74.65 and 78%, respectively. The proposed model performed better than previous other models. Conclusions: The proposed model appears to be effective in analyzing arousal and valence; specifically, it is expected that affective analysis using physiological signals based on LSTM will be possible without manual feature extraction. In a future study, the classification model will be adopted in mental healthcare management systems.

In this study, automatic detection of stator winding inter-turn short circuit fault (SWISCFs) in surface-mounted permanent magnet synchronous motors (SPMSMs) and automatic classification of fault severity via a pattern recognition system (PRS) are presented. In the case of a stator short circuit fault, performance losses become an important issue for SPMSMs. To detect stator winding short circuit faults automatically and to estimate the severity of the fault, an artificial neural network (ANN)-based PRS was used. It was found that the amplitude of the third harmonic of the current was the most distinctive characteristic for detecting the short circuit fault ratio of the SPMSM. To validate the proposed method, both simulation results and experimental results are presented.

기존 안드로이드 단말기에서 제공되는 문자기능은 단순히 수신만 가능한 것에 그쳐 사용자가 원하는 정보만 얻음에 있어 다소 불편함이 있었다. 기존 불편함을 보완하고자 정규표현식을 이용한 단문 메시지 파싱 기능을 구현하여 원하는 정보를 분류해서 볼 수 있게 구현하였다.

기존 안드로이드 단말기에서 제공되는 문자기능은 단순히 수신만 가능한 것에 그쳐 사용자가 원하는 정보만 얻음에 있어 다소 불편함이 있었다. 기존 불편함을 보완하고자 정규표현식을 이용한 단문 메시지 파싱 기능을 구현하여 원하는 정보를 분류해서 볼 수 있게 구현하였다.

현대 사회에서는 짧은 길이의 문장으로 이루어진 문서가 증가하고 있어, 이를 효과적으로 관리하고자 카테고리별로 분류하는 작업이 중요한 과제로 부각 되고 있으며, 이에 따라 효율적이고 성능이 좋은 분류시스템의 필요성이 대두되고 있다. 본 연구에서는 기존에 발표된 연구에서 주로 대상으로 삼았던 문서들보다 훨씬 짧은 길이의 1~2문장으로 구성된 문서를 대상으로 하며, 이러한 문서들은 다양한 주제와 형식으로 구성되어 있다. 이러한 문서들은 문맥의 부족, 데이터 불균형 등의 특성이 있기에 발생하는 분류의 어려움이 존재한다. 따라서 해당 특성을 고려하여 다양한 기계학습 알고리즘을 조사하고, 다양한 기계학습 알고리즘에 대한 학습을 진행하여 분류성능을 비교한다. 본 연구에서는 로지스틱 회귀, Support Vector Machine, Random Forest, XGBoost, LSTM, BERT 기법을 사용하였다. 이후 분류성능이 뛰어난 알고리즘을 바탕으로 앙상블 기법을 적용하여 단문 분류에서의 정확도를 5~6% 향상했다.

In modern society, the number of documents composed of short sentences is increasing, and classifying them by category is emerging as an important task to effectively manage them, and the need for an efficient and high-performance classification system is emerging. This study targets much shorter documents than those that were mainly targeted in previously published studies, and these documents are composed of various topics and formats. These documents have difficulty in classification because they have characteristics such as lack of context and data imbalance. Therefore, various machine learning algorithms are investigated in consideration of their characteristics, and the classification performance is compared by learning about various machine learning algorithms. Logistic Regression, Support Vector Machine, Random Forest, XGBoost, LSTM and BERT techniques were used in this study. Afterwards, the accuracy in short text classification was improved by 5-6% by applying the ensemble technique based on an algorithm with excellent classification performance.

Pupil size variation can not be controlled intentionally by the user and includes various features such as the blinking frequency and the duration of a blink, so it is suitable for understanding the user's emotional state. In addition, an ocular feature based emotion classification method should be studied for virtual and augmented reality, which is expected to be applied to various fields. In this paper, we propose a novel emotion classification based on CNN with pupil size variation signals which include not only various ocular feature information but also time information. As a result, compared to previous studies using the same database, the proposed method showed improved results of 5.99% and 12.98% respectively from arousal and valence emotion classification.

As smart devices and social network services (SNSs) become increasingly pervasive, individuals produce large amounts of data in real time. Accordingly, studies on unstructured data analysis are actively being conducted to solve the resultant problem of information overload and to facilitate effective data processing. Many such studies are conducted for filtering inappropriate information. In this paper, a feature-weighting method considering SNS-message features is proposed for the classification of short text messages generated on SNSs, using Markov logic networks for category inference. The performance of the proposed method is verified through a comparison with an existing frequency-based classification methods.

The purpose of this study is to understand the characteristics of body types of adult woman under 155cm in height. To do this, direct measurement data showed that only 660 women aged 20~69 years old and under 155cm tall were classified, and their body type characteristics were analyzed. There are 51 body measurement items used in the analysis. The results of the study are as follows: As a result of the comparison of the body size of all adult women and women under 155cm, it was seen that short women were smaller than the all women, and were flattened with a slightly plump and bendless body. Factor analysis was performed to classify the body type of short women. Six factors including body size, body height, shoulder length and width, waist back length, hip length, and shoulder angle were measured. The results of cluster analysis, using factor scores, were classified into three types. Type 1 is taller among short women and slightly slimmer, and is an A-shape. Type 2 has the largest BMI and was the middle stature among short women It is the most obese body type and is an H-shape. Type 3 is short and plump, with a poor hip and is a Y-shape. As a result of crossover analysis, there was a difference in the body type characteristics of each age group. After all, the body type of the short adult women under 155cm was clearly different from those of all adult women, there was also a difference in the body type among short adult women.

소매유통 기업의 수요관리에서 수요예측의 중요성은 증가하고 있다. 시계열 데이터를 활용한 수요예측의 정확도를 높이기 위해서는 제품의 수요 특성을 이해하고 이에 따라 예측 모형을 구성하는 것이 필요하다. 다양한 종류의 제품을 판매하는 소매유통 기업은 개별 상품별로 시계열 수요예측 모형을 구성하기 어렵기 때문에 시계열 수요 특성이 동일하거나 유사한 상품군을 분류하고 상품군별로 최적의 예측 모형을 구성하는 방법이 요구된다. 본 연구에서는 수요의 유형 분류를 위하여 전통적인 SBC 방법과 같이 시계열 수요 데이터의 일부 통계량을 활용한 정적 수요 분류 방법의 한계를 보완하기 위하여 시계열 데이터를 직접 활용한 시계열 군집분석의 분류 성능을 평가하고자 한다. M5 competition에서 제공하는 총 1,246개 개별 상품의 2년치 판매 데이터를 대상으로 수치실험을 수행하였다. 수치실험 결과 다음과 같은 결과를 확인하였다. 첫째, Shape 기반 방법과 계층적 방법을 각각 유사도 측정 방법과 군집 구성 방법으로 사용한 시계열 군집분석 방법의 성능이 가장 우수하였다. 둘째, 전통적인 SBC 분류 방법은 직관적이고 계산이 용이한 장점이 있지만 시계열 수요 데이터의 특성을 반영하지 못하는 한계를 확인하였다. 셋째, 가용 데이터의 양과 데이터의 세분성 수준과 관계없이 시계열 군집분석 방법이 SBC 수요분류 방법 대비 우수한 성능을 보였다.

The importance of demand forecasting is growing in retail industry. To improve the accuracy of demand forecasting using time series data, it is crucial to understand the demand characteristics and design forecasting models accordingly. Retail companies that sell a wide variety of products find it difficult to develop individual time series forecasting models for each product. Therefore, there is a need to classify products with similar or identical time series demand characteristics and develop optimal forecasting models for each product group. This study aims to evaluate the classification performance of time series clustering method, which directly utilizes time series data, to complement the limitations of traditional static demand classification methods using summary statistics such as the traditional SBC method. Numerical experiments were conducted using sales data for 1,246 individual products over two years, provided by the M5 competition. The results of the numerical experiments are as follows: First, the performance of time series clustering method, using Shape-based similarity measure and hierarchical clustering method, is the best. Second, while the traditional SBC classification method has the advantages of being intuitive and computationally simple, it was found to have limitations in reflecting the characteristics of time series demand. Third, regardless of the amount of available data or the level of data granularity, the time series clustering method showed superior performance compared to the SBC demand classification method.

음성인식처리 분야에서 배경잡음으로 인하여 음성입력이 배경잡음으로 잘못 판단되는 원인이 되어 음성인식율의 저하를 초래한다. 이러한 종류의 잡음대책은 단순하지 않으므로 보다 고도한 잡음처리기술이 필요하게 된다. 따라서 본 논문에서는 잡음환경 중에서 정상적인 배경잡음 혹은 비정상적인 배경잡음과 지속 시간이 짧은 음성을 구별하는 알고리즘에 대하여 기술한다. 본 알고리즘은 다른 종류의 잡음과 음성을 구별하는 중요한 수단으로서 개량된 음성의 특징파리미터를 사용한다. 다음으로 다층퍼셉트론 네트워크에 의하여 잡음의 종류를 추정하는 알고리즘에 대해서 기술한다. 본 실험에서는 잡음과 음성이 구별이 가능하도록 실험적으로 확인하였다.

In the area of the speech recognition processing, background noises are caused the incorrect response to the speech input, therefore the speech recognition rates are decreased by the background noises. Accordingly, a more high level noise processing techniques are required since these kinds of noise countermeasures are not simple. Therefore, this paper proposes an algorithm to distinguish between the stationary background noises or non-stationary background noises and the speech signal having short time duration in the noisy environments. The proposed algorithm uses the characteristic parameter of the improved speech signal as an important measure in order to distinguish different types of the background noises and the speech signals. Next, this algorithm estimates various kinds of the background noises using a multi-layer perceptron neural network. In this experiment, it was experimentally clear the estimation of the background noises and the speech signals.