• KIISE Transactions on Computing Practices
• /
• v.23 no.5
• /
• pp.299-303
• /
• 2017

By piling up hidden layers in artificial neural networks, deep learning is delivering outstanding performances for high-level abstraction problems such as object/speech recognition and natural language processing. Alternatively, deep-learning users often struggle with the tremendous amounts of time and resources that are required to train deep neural networks. To alleviate this computational challenge, many approaches have been proposed in a diversity of areas. In this work, two of the existing Apache Spark-based acceleration frameworks for deep learning (SparkNet and DeepSpark) are compared and analyzed in terms of the training accuracy and the time demands. In the authors' experiments with the CIFAR-10 and CIFAR-100 benchmark datasets, SparkNet showed a more stable convergence behavior than DeepSpark; but in terms of the training accuracy, DeepSpark delivered a higher classification accuracy of approximately 15%. For some of the cases, DeepSpark also outperformed the sequential implementation running on a single machine in terms of both the accuracy and the running time.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2018.05a
• /
• pp.466-467
• /
• 2018

Recently, artificial intelligence has been applied to various fields such as image recognition, image recognition speech recognition, and natural language processing, and interest in Deep Learning technology is increasing. Many researches on Convolutional Neural Network(CNN), which is one of the most representative algorithms among Deep Learning, have strong advantages in image recognition and classification and are widely used in various fields. In this paper, we propose a new network structure that transforms the general CNN structure. A typical CNN structure consists of a convolution layer, ReLU layer, and a pooling layer. Therefore in this paper, We intend to construct a new network by adding fully connected layer inside a general CNN structure. This modification is intended to increase the learning and accuracy of the convoluted image by including the generalization which is an advantage of the neural network.

• The Journal of Information Systems
• /
• v.29 no.3
• /
• pp.237-251
• /
• 2020

Purpose The research was studied the hierarchical Hangul emotion index by organizing all the emotions which SNS users are thinking. As a preliminary study by the researcher, the English-based Plutchick (1980)'s emotional standard was reinterpreted in Korean, and a hashtag with implicit meaning on SNS was studied. To build a multidimensional emotion dictionary and classify three-dimensional emotions, an emotion seed was selected for the composition of seven emotion sets, and an emotion word dictionary was constructed by collecting SNS hashtags derived from each emotion seed. We also want to explore the priority of each Hangul emotion index. Design/methodology/approach In the process of transforming the matrix through the vector process of words constituting the sentence, weights were extracted using TF-IDF (Term Frequency Inverse Document Frequency), and the dimension reduction technique of the matrix in the emotion set was NMF (Nonnegative Matrix Factorization) algorithm. The emotional dimension was solved by using the characteristic value of the emotional word. The cosine distance algorithm was used to measure the distance between vectors by measuring the similarity of emotion words in the emotion set. Findings Customer needs analysis is a force to read changes in emotions, and Korean emotion word research is the customer's needs. In addition, the ranking of the emotion words within the emotion set will be a special criterion for reading the depth of the emotion. The sentiment index study of this research believes that by providing companies with effective information for emotional marketing, new business opportunities will be expanded and valued. In addition, if the emotion dictionary is eventually connected to the emotional DNA of the product, it will be possible to define the "emotional DNA", which is a set of emotions that the product should have.

• Journal of the Korea Convergence Society
• /
• v.9 no.7
• /
• pp.25-31
• /
• 2018

Online news has replaced the traditional newspaper and has brought about a profound transformation in the way we access and share information. News websites have had the ability for users to post comments for quite some time, and some have also begun to crowdsource reactions to news articles. The field of sentiment analysis seeks to computationally model the emotions and reactions experienced when presented with text. In this work, we analyze more than 100,000 news articles over ten categories with five user-generated emotional annotations to determine whether or not these reactions have a mathematical correlation to the news body text and propose a simple sentiment analysis algorithm that requires minimal preprocessing and no machine learning. We show that it is effective even for a morphologically complex language like Korean.

• Journal of the Korea Convergence Society
• /
• v.9 no.10
• /
• pp.45-53
• /
• 2018

Recently, studies about bilingual word embedding have been gaining much attention. However, bilingual word embedding with Korean is not actively pursued due to the difficulty in obtaining a sizable, high quality corpus. Local embeddings that can be applied to specific domains are relatively rare. Additionally, multi-word vocabulary is problematic due to the lack of one-to-one word-level correspondence in translation pairs. In this paper, we crawl 868,163 paragraphs from a Korean-English law corpus and propose three mapping strategies for word embedding. These strategies address the aforementioned issues including multi-word translation and improve translation pair quality on paragraph-aligned data. We demonstrate a twofold increase in translation pair quality compared to the global bilingual word embedding baseline.

• Journal of KIISE
• /
• v.42 no.11
• /
• pp.1397-1403
• /
• 2015

A named entity(NE) dictionary is an important resource for the performance of NE recognition. However, it is not easy to construct a NE dictionary manually since human annotation is time consuming and labor-intensive. To save construction time and reduce human labor, we propose a semi-automatic system for the construction of a NE dictionary. The proposed system constructs a pseudo-document with Wiki-categories per NE class by using an active learning technique. Then, it calculates similarities between Wiki entries and pseudo-documents using the BM25 model, a well-known information retrieval model. Finally, it classifies each Wiki entry into NE classes based on similarities. In experiments with three different types of NE class sets, the proposed system showed high performance(macro-average F1-score of 0.9028 and micro-average F1-score 0.9554).

• Journal of KIISE
• /
• v.44 no.11
• /
• pp.1138-1148
• /
• 2017

Topic extraction is a technology that automatically extracts a set of topics from a set of documents, and this has been a major research topic in the area of natural language processing. Representative topic extraction methods include Latent Dirichlet Allocation (LDA) and word clustering-based methods. However, there are problems with these methods, such as repeated topics and mixed topics. The problem of repeated topics is one in which a specific topic is extracted as several topics, while the problem of mixed topic is one in which several topics are mixed in a single extracted topic. To solve these problems, this study proposes a method to extract topics using an LDA that is robust against the problem of repeated topic, going through the steps of separating and merging the topics using the similarity between words to correct the extracted topics. As a result of the experiment, the proposed method showed better performance than the conventional LDA method.

• The Korean Journal of Archival Studies
• /
• no.67
• /
• pp.99-138
• /
• 2021

In the era of big data, analyzing not only structured data but also unstructured data is emerging as an important task. Official documents produced by government agencies are also subject to big data analysis as large text-based unstructured data. From the perspective of internal work efficiency, knowledge management, records management, etc, it is necessary to analyze big data of public documents to derive useful implications. However, since many of the public documents currently held by public institutions are not in open format, a pre-processing process of extracting text from a bitstream is required for big data analysis. In addition, since contextual metadata is not sufficiently stored in the document file, separate efforts to secure metadata are required for high-quality analysis. In conclusion, the current official documents have a low level of machine readability, so big data analysis becomes expensive.

• The Korean Journal of Applied Statistics
• /
• v.34 no.3
• /
• pp.389-399
• /
• 2021

In recent years, word embedding has been a popular field of natural language processing research and a skip-gram has become one successful word embedding method. It assigns a word embedding vector to each word using contexts, which provides an effective way to analyze text data. However, due to the limitation of vector space model, primary word embedding methods assume that every word only have a single meaning. As one faces multi-sense words, that is, words with more than one meaning, in reality, Neelakantan (2014) proposed a multi-sense skip-gram (MSSG) to find embedding vectors corresponding to the each senses of a multi-sense word using a clustering method. In this paper, we propose a modified method of the MSSG to improve statistical accuracy. Moreover, we propose a data-adaptive choice of the number of clusters, that is, the number of meanings for a multi-sense word. Some numerical evidence is given by conducting real data-based simulations.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.26 no.5
• /
• pp.662-667
• /
• 2022

Semantic role labeling task(SRL) is to extract predicate and arguments such as agent, patient, place, time. In the previously SRL task studies, a pipeline method extracting linguistic features of sentence has been proposed, but in this method, errors of each extraction work in the pipeline affect semantic role labeling performance. Therefore, methods using End-to-End neural network model have recently been proposed. In this paper, we propose a neural network model using the Biaffine Average Attention model for SRL task. The proposed model consists of a structure that can focus on the entire sentence information regardless of the distance between the predicate in the sentence and the arguments, instead of LSTM model that uses the surrounding information for prediction of a specific token proposed in the previous studies. For evaluation, we used F1 scores to compare two models based BERT model that proposed in existing studies using F1 scores, and found that 76.21% performance was higher than comparison models.