• Journal of Intelligence and Information Systems
• /
• v.22 no.2
• /
• pp.127-142
• /
• 2016

Deep learning model is a kind of neural networks that allows multiple hidden layers. There are various deep learning architectures such as convolutional neural networks, deep belief networks and recurrent neural networks. Those have been applied to fields like computer vision, automatic speech recognition, natural language processing, audio recognition and bioinformatics where they have been shown to produce state-of-the-art results on various tasks. Among those architectures, convolutional neural networks and recurrent neural networks are classified as the supervised learning model. And in recent years, those supervised learning models have gained more popularity than unsupervised learning models such as deep belief networks, because supervised learning models have shown fashionable applications in such fields mentioned above. Deep learning models can be trained with backpropagation algorithm. Backpropagation is an abbreviation for "backward propagation of errors" and a common method of training artificial neural networks used in conjunction with an optimization method such as gradient descent. The method calculates the gradient of an error function with respect to all the weights in the network. The gradient is fed to the optimization method which in turn uses it to update the weights, in an attempt to minimize the error function. Convolutional neural networks use a special architecture which is particularly well-adapted to classify images. Using this architecture makes convolutional networks fast to train. This, in turn, helps us train deep, muti-layer networks, which are very good at classifying images. These days, deep convolutional networks are used in most neural networks for image recognition. Convolutional neural networks use three basic ideas: local receptive fields, shared weights, and pooling. By local receptive fields, we mean that each neuron in the first(or any) hidden layer will be connected to a small region of the input(or previous layer's) neurons. Shared weights mean that we're going to use the same weights and bias for each of the local receptive field. This means that all the neurons in the hidden layer detect exactly the same feature, just at different locations in the input image. In addition to the convolutional layers just described, convolutional neural networks also contain pooling layers. Pooling layers are usually used immediately after convolutional layers. What the pooling layers do is to simplify the information in the output from the convolutional layer. Recent convolutional network architectures have 10 to 20 hidden layers and billions of connections between units. Training deep learning networks has taken weeks several years ago, but thanks to progress in GPU and algorithm enhancement, training time has reduced to several hours. Neural networks with time-varying behavior are known as recurrent neural networks or RNNs. A recurrent neural network is a class of artificial neural network where connections between units form a directed cycle. This creates an internal state of the network which allows it to exhibit dynamic temporal behavior. Unlike feedforward neural networks, RNNs can use their internal memory to process arbitrary sequences of inputs. Early RNN models turned out to be very difficult to train, harder even than deep feedforward networks. The reason is the unstable gradient problem such as vanishing gradient and exploding gradient. The gradient can get smaller and smaller as it is propagated back through layers. This makes learning in early layers extremely slow. The problem actually gets worse in RNNs, since gradients aren't just propagated backward through layers, they're propagated backward through time. If the network runs for a long time, that can make the gradient extremely unstable and hard to learn from. It has been possible to incorporate an idea known as long short-term memory units (LSTMs) into RNNs. LSTMs make it much easier to get good results when training RNNs, and many recent papers make use of LSTMs or related ideas.

• Journal of Intelligence and Information Systems
• /
• v.24 no.2
• /
• pp.59-83
• /
• 2018

With the increasing importance of sentiment analysis to grasp the needs of customers and the public, various types of deep learning models have been actively applied to English texts. In the sentiment analysis of English texts by deep learning, natural language sentences included in training and test datasets are usually converted into sequences of word vectors before being entered into the deep learning models. In this case, word vectors generally refer to vector representations of words obtained through splitting a sentence by space characters. There are several ways to derive word vectors, one of which is Word2Vec used for producing the 300 dimensional Google word vectors from about 100 billion words of Google News data. They have been widely used in the studies of sentiment analysis of reviews from various fields such as restaurants, movies, laptops, cameras, etc. Unlike English, morpheme plays an essential role in sentiment analysis and sentence structure analysis in Korean, which is a typical agglutinative language with developed postpositions and endings. A morpheme can be defined as the smallest meaningful unit of a language, and a word consists of one or more morphemes. For example, for a word '예쁘고', the morphemes are '예쁘(= adjective)' and '고(=connective ending)'. Reflecting the significance of Korean morphemes, it seems reasonable to adopt the morphemes as a basic unit in Korean sentiment analysis. Therefore, in this study, we use 'morpheme vector' as an input to a deep learning model rather than 'word vector' which is mainly used in English text. The morpheme vector refers to a vector representation for the morpheme and can be derived by applying an existent word vector derivation mechanism to the sentences divided into constituent morphemes. By the way, here come some questions as follows. What is the desirable range of POS(Part-Of-Speech) tags when deriving morpheme vectors for improving the classification accuracy of a deep learning model? Is it proper to apply a typical word vector model which primarily relies on the form of words to Korean with a high homonym ratio? Will the text preprocessing such as correcting spelling or spacing errors affect the classification accuracy, especially when drawing morpheme vectors from Korean product reviews with a lot of grammatical mistakes and variations? We seek to find empirical answers to these fundamental issues, which may be encountered first when applying various deep learning models to Korean texts. As a starting point, we summarized these issues as three central research questions as follows. First, which is better effective, to use morpheme vectors from grammatically correct texts of other domain than the analysis target, or to use morpheme vectors from considerably ungrammatical texts of the same domain, as the initial input of a deep learning model? Second, what is an appropriate morpheme vector derivation method for Korean regarding the range of POS tags, homonym, text preprocessing, minimum frequency? Third, can we get a satisfactory level of classification accuracy when applying deep learning to Korean sentiment analysis? As an approach to these research questions, we generate various types of morpheme vectors reflecting the research questions and then compare the classification accuracy through a non-static CNN(Convolutional Neural Network) model taking in the morpheme vectors. As for training and test datasets, Naver Shopping's 17,260 cosmetics product reviews are used. To derive morpheme vectors, we use data from the same domain as the target one and data from other domain; Naver shopping's about 2 million cosmetics product reviews and 520,000 Naver News data arguably corresponding to Google's News data. The six primary sets of morpheme vectors constructed in this study differ in terms of the following three criteria. First, they come from two types of data source; Naver news of high grammatical correctness and Naver shopping's cosmetics product reviews of low grammatical correctness. Second, they are distinguished in the degree of data preprocessing, namely, only splitting sentences or up to additional spelling and spacing corrections after sentence separation. Third, they vary concerning the form of input fed into a word vector model; whether the morphemes themselves are entered into a word vector model or with their POS tags attached. The morpheme vectors further vary depending on the consideration range of POS tags, the minimum frequency of morphemes included, and the random initialization range. All morpheme vectors are derived through CBOW(Continuous Bag-Of-Words) model with the context window 5 and the vector dimension 300. It seems that utilizing the same domain text even with a lower degree of grammatical correctness, performing spelling and spacing corrections as well as sentence splitting, and incorporating morphemes of any POS tags including incomprehensible category lead to the better classification accuracy. The POS tag attachment, which is devised for the high proportion of homonyms in Korean, and the minimum frequency standard for the morpheme to be included seem not to have any definite influence on the classification accuracy.

• The KIPS Transactions:PartB
• /
• v.19B no.3
• /
• pp.195-200
• /
• 2012

In Korean, spelling change in various forms must be recovered into base forms in morphological analysis as well as part-of-speech (POS) tagging is difficult without morphological analysis because Korean is agglutinative. This is one of notorious problems in Korean morphological analysis and has been solved by morpheme recovery rules, which generate morphological ambiguity resolved by POS tagging. In this paper, we propose a morpheme recovery scheme based on machine learning methods like Na$\ddot{i}$ve Bayes models. Input features of the models are the surrounding context of the syllable which the spelling change is occurred and categories of the models are the recovered syllables. The POS tagging system with the proposed model has demonstrated the $F_1$-score of 97.5% for the ETRI tree-tagged corpus. Thus it can be decided that the proposed model is very useful to handle morpheme recovery in Korean.

• Journal of the Korean Institute of Intelligent Systems
• /
• v.12 no.6
• /
• pp.517-523
• /
• 2002

In this paper, we propose system that apply VIBL method to add speech recognition to LIBL method based on human s studying method to use natural language to steering system of ship, MERCS and winch appliances and use VIBL method to alternate process that linguistic instruction such as officer s steering instruction is achieved via ableman and control steering gear, MERCS and winch appliances. By specific method of study, ableman s suitable steering manufacturing model embodies intelligent steering gear controlling system that embody and language direction base studying method to present proper meaning element and evaluation rule to steering system of ship apply and respond more efficiently on voice instruction of commander using fuzzy inference rule. Also we embody system that recognize voice direction of commander and control MERCS and winch appliances. We embodied steering manufacturing model based on ableman s experience and presented rudder angle for intelligent steering system, compass bearing arrival time, evaluation rule to propose meaning element of stationary state and correct steerman manufacturing model rule using technique to recognize voice instruction of commander and change to text and fuzzy inference. Also we apply VIBL method to speech recognition ship control simulator and confirmed the effectiveness.

• Annual Conference on Human and Language Technology
• /
• 1995.10a
• /
• pp.169-174
• /
• 1995

품사 태깅은 형태소 분석 이후 발생한 모호성을 제거하는 것으로, 통계적 방법과 규칙에 기 반한 방법이 널리 사용되고 있다. 하지만, 이들 방법론에는 각기 한계점을 지니고 있다. 통계적인 방법인 은닉 마코프 모델(Hidden Markov Model)은 유연성(flexibility)을 지니지만, 교착어(agglutinative language)인 한국어에 있어서 제한된 윈도우로 인하여, 중의성 해결의 실마리가 되는 어휘나 품사별 제대로 참조하지 못하는 경우가 있다. 반면, 규칙에 기반한 방법은 차체가 품사에 영향을 받으므로 인하여, 새로운 태그집합(tagset)이나 언어에 대하여 유연성이나 정확성을 제공해 주지 못한다. 이러한 각기 서로 다른 방법론의 한계를 극복하기 위하여, 본 논문에서는 통계와 규칙을 통합한 한국어 태깅 모델을 제안한다. 즉 통계적 학습을 통한 통계 모델이후에 2차적으로 규칙을 자동학습 하게 하여, 통계모델이 다루지 못하는 범위의 규칙을 생성하게 된다. 이처럼 2단계의 통계와 규칙의 자동 학습단계를 거치게 됨으로써, 두개 모델의 단점을 보강한 높은 정확도를 가지는 한국어 태거를 개발할 수 있게 하였다.

• Proceedings of the KSPS conference
• /
• 2006.11a
• /
• pp.14-18
• /
• 2006

The role of dialogue manager is to select proper actions based on observed environment and inferred user intention. This paper presents stochastic model for dialogue manager based on Markov decision process. To build a mixed initiative dialogue manager, we used accumulated user utterance, previous act of dialogue manager, and domain dependent knowledge as the input to the MDP. We also used dialogue corpus to train the automatically optimized policy of MDP with reinforcement learning algorithm. The states which have unique and intuitive actions were removed from the design of MDP by using the domain knowledge. The design of dialogue manager included the usage of natural language understanding and response generator to build short message based remote control of home networked appliances.

• Journal of IKEEE
• /
• v.22 no.1
• /
• pp.156-161
• /
• 2018

Recently, as the population rapidly increases, food shortages and waste are emerging serious problem. In order to solve this problem, various countries and enterprises are trying research and product development such as a study of consumers' purchasing patterns of food and a development of smart refrigerator using IoT technology. However, the smart refrigerators which currently sold have high price issue and another waste due to malfunction and breakage by complicated configurations. In this paper, we proposed a low-cost smart refrigerator system based on IoT for solving the problem and efficient management of ingredients. The system recognizes and registers ingredients through QR code, image recognition, and speech recognition, and can provide various services of the smart refrigerator. In order to improve an accuracy of image recognition, we used a model using a deep learning algorithm and proved that it is possible to register ingredients accurately.

• Journal of the Korea Convergence Society
• /
• v.11 no.6
• /
• pp.203-208
• /
• 2020

This study examined whether Mild Cognitive Impairment (MCI) is related to the reduction of specific memory among linguistic memory and visuospatial memory, and to identify the most predictive index for discriminating MCI from normal elderly. The subjects were analyzed for 189 elderly (103 healthy elderly, 86 MCI). The verbal memory was used by the Seoul Verbal Learning Test. visuospatial memory was measured using the Rey Complex Figure Test. As a result of multiple logistic regression, verbal memory and visuospatial memory showed significant predictive performance in discriminating MCI from normal elderly. On the other hand, when all the confounding variables were corrected, including the results of each memory test, the predictive power was significant in distinguishing MCI from normal aging only in the immediate recall of verbal memory, and the predictive power was not significant in the immediate recall of visuospatial memory. This result suggests that delayed recall of visuospatial memory and immediate recall of verbal memory are the best combinations to discriminate memory ability of MCI.

• Journal of the Korea Institute of Information Security & Cryptology
• /
• v.32 no.1
• /
• pp.89-98
• /
• 2022

As deep learning technology was applied to various fields, research on adversarial attack techniques, a security problem of deep learning models, was actively studied. adversarial attacks have been mainly studied in the field of images. Recently, they have even developed a complete decision-based attack technique that can attack with just the classification results of the model. However, in the case of the audio field, research is relatively slow. In this paper, we applied several decision-based attack techniques to the audio field and improved state-of-the-art attack techniques. State-of-the-art decision-attack techniques have the disadvantage of requiring many queries for gradient approximation. In this paper, we improve query efficiency by proposing a method of reducing the vector search space required for gradient approximation. Experimental results showed that the attack success rate was increased by 50%, and the difference between original audio and adversarial examples was reduced by 75%, proving that our method could generate adversarial examples with smaller noise.

• Journal of Digital Convergence
• /
• v.20 no.5
• /
• pp.53-59
• /
• 2022

The purpose of this study is to examine the factors related to learning through analysis of domestic research trends in literacy and to present the direction of literacy education. Research papers from 1993 to February 2022 were collected using RISS. 'Literacy' and 'Education' were used as search keywords, and 200 papers were selected for analysis. As a result of analysis using keyword network analysis, 118 keywords appeared at least three times out of a total of 810 keywords. The order of the keywords with the highest frequency is 'digital literacy', 'media literacy', and 'elementary school'. The following direction was suggested through the analysis results. First, it is required to establish an online teaching and learning resource platform and link it with education policy. Second, it is necessary to set literacy competencies and seek ways to improve competencies. Third, a digital-based convergence education model should be developed. This study is meaningful in that it analyzed the most recent literacy studies and suggested the direction of literacy education.