• Journal of Information Processing Systems
• /
• v.17 no.5
• /
• pp.1020-1033
• /
• 2021

Nowadays, online or mobile social network services (SNS) are very popular and widely spread in our society and daily lives to instantly share, disseminate, and search information. In particular, SNS such as YouTube, Flickr, Facebook, and Amazon allow users to upload billions of images or videos and also provide a number of multimedia information to users. Information retrieval in multimedia-rich SNS is very useful but challenging task. Content-based media retrieval (CBMR) is the process of obtaining the relevant image or video objects for a given query from a collection of information sources. However, CBMR suffers from the dimensionality curse due to inherent high dimensionality features of media data. This paper investigates the effectiveness of the kernel trick in CBMR, specifically, the kernel principal component analysis (KPCA) for dimensionality reduction. KPCA is a nonlinear extension of linear principal component analysis (LPCA) to discovering nonlinear embeddings using the kernel trick. The fundamental idea of KPCA is mapping the input data into a highdimensional feature space through a nonlinear kernel function and then computing the principal components on that mapped space. This paper investigates the potential of KPCA in CBMR for feature extraction or dimensionality reduction. Using the Gaussian kernel in our experiments, we compute the principal components of an image dataset in the transformed space and then we use them as new feature dimensions for the image dataset. Moreover, KPCA can be applied to other many domains including CBMR, where LPCA has been used to extract features and where the nonlinear extension would be effective. Our results from extensive experiments demonstrate that the potential of KPCA is very encouraging compared with LPCA in CBMR.

• International Journal of Computer Science & Network Security
• /
• v.21 no.9
• /
• pp.281-291
• /
• 2021

The Islamic jurisprudential legal system represents an essential component of the Islamic religion, that governs many aspects of Muslims' daily lives. This creates many questions that require interpretations by qualified specialists, or Muftis according to the main sources of legislation in Islam. The Islamic jurisprudence is usually classified into branches, according to which the questions can be categorized and classified. Such categorization has many applications in automated question-answering systems, and in manual systems in routing the questions to a specialized Mufti to answer specific topics. In this work we tackle the problem of automatic categorisation of Islamic jurisprudential legal questions using deep learning techniques. In this paper, we build a hierarchical deep learning model that first extracts the question text features at two levels: word and sentence representation, followed by a text classifier that acts upon the question representation. To evaluate our model, we build and release the largest publicly available dataset of Islamic questions and answers, along with their topics, for 52 topic categories. We evaluate different state-of-the art deep learning models, both for word and sentence embeddings, comparing recurrent and transformer-based techniques, and performing extensive ablation studies to show the effect of each model choice. Our hierarchical model is based on pre-trained models, taking advantage of the recent advancement of transfer learning techniques, focused on Arabic language.

• Journal of Information Science Theory and Practice
• /
• v.9 no.2
• /
• pp.1-17
• /
• 2021

Pseudo relevance feedback (PRF) is a powerful query expansion (QE) technique that prepares queries using the top k pseudorelevant documents and choosing expansion elements. Traditional PRF frameworks have robustly handled vocabulary mismatch corresponding to user queries and pertinent documents; nevertheless, expansion elements are chosen, disregarding similarity to the original query's elements. Word embedding (WE) schemes comprise techniques of significant interest concerning QE, that falls within the information retrieval domain. Deep averaging networks (DANs) defines a framework relying on average word presence passed through multiple linear layers. The complete query is understandably represented using the average vector comprising the query terms. The vector may be employed for determining expansion elements pertinent to the entire query. In this study, we suggest a DANs-based technique that augments PRF frameworks by integrating WE similarities to facilitate Arabic information retrieval. The technique is based on the fundamental that the top pseudo-relevant document set is assessed to determine candidate element distribution and select expansion terms appropriately, considering their similarity to the average vector representing the initial query elements. The Word2Vec model is selected for executing the experiments on a standard Arabic TREC 2001/2002 set. The majority of the evaluations indicate that the PRF implementation in the present study offers a significant performance improvement compared to that of the baseline PRF frameworks.

• Journal of Internet of Things and Convergence
• /
• v.8 no.5
• /
• pp.17-25
• /
• 2022

Recently the massive amount of data has been generated because of the number of edge devices increases. And especially, the number of raw unstructured HTML documents has been increased. Therefore, MRC(Machine Reading Comprehension) in which a natural language processing model finds the important information within an HTML document is becoming more important. In this paper, we propose HTDE(HTML Tag Depth Embedding Method), which allows the BERT to train the depth of the HTML document structure. HTDE makes a tag stack from the HTML document for each input token in the BERT and then extracts the depth information. After that, we add a HTML embedding layer that takes the depth of the token as input to the step of input embedding of BERT. Since tokenization using HTDE identifies the HTML document structures through the relationship of surrounding tokens, HTDE improves the accuracy of BERT for HTML documents. Finally, we demonstrated that the proposed idea showing the higher accuracy compared than the accuracy using the conventional embedding of BERT.

• Journal of the Korea Computer Graphics Society
• /
• v.29 no.4
• /
• pp.7-16
• /
• 2023

For decades, creating a desired locomotive motion in a goal-oriented manner has been a challenge in character animation. Data-driven methods using generative models have demonstrated efficient ways of predicting long sequences of motions without the need for explicit conditioning. While these methods produce high-quality long-term motions, they can be limited when it comes to synthesizing motion for challenging novel scenarios, such as punching a random target. A state-of-the-art solution to overcome this limitation is by using a GAN Discriminator to imitate motion data clips and incorporating reinforcement learning to compose goal-oriented motions. In this paper, our research aims to create characters performing combat sports such as boxing, using a novel reward design in conjunction with existing GAN-based approaches. We experimentally demonstrate that both the Adversarial Motion Prior [3] and Adversarial Skill Embeddings [4] methods are capable of generating viable motions for a character punching a random target, even in the absence of mocap data that specifically captures the transition between punching and locomotion. Also, with a single learned policy, multiple task controllers can be constructed through the TimeChamber framework.

• The Journal of The Korea Institute of Intelligent Transport Systems
• /
• v.22 no.6
• /
• pp.17-31
• /
• 2023

An activity-based model requires detailed population information to model individual travel behavior in a disaggregated manner. The recent innovative approach developed deep generative models with novel regularization terms that improves fidelity and diversity for population synthesis. Since the method relies on measuring the distance between distribution boundaries of the sample data and the generated sample, it is crucial to obtain well-defined continuous representation from the discretized dataset. Therefore, we propose an improved entity embedding models to enhance the performance of the regularization terms, which indirectly supports the synthesis in terms of feasible and diverse populations. Our results show a 28.87% improvement in the F1 score compared to the baseline method.

• Journal of the Korean Society for Library and Information Science
• /
• v.51 no.4
• /
• pp.99-120
• /
• 2017

In order to develop a semiautomatic support system that allows researchers concerned to efficiently analyze the technical trends for the ever-growing industry and market. This paper introduces a couple of Korean sentence generation models that can automatically generate definitional statements as well as descriptions of technical terms and concepts. The proposed models are based on a deep learning model called LSTM (Long Sort-Term Memory) capable of effectively labeling textual sequences by taking into account the contextual relations of each item in the sequences. Our models take technical terms as inputs and can generate a broad range of heterogeneous textual descriptions that explain the concept of the terms. In the experiments using large-scale training collections, we confirmed that more accurate and reasonable sentences can be generated by CHAR-CNN-LSTM model that is a word-based LSTM exploiting character embeddings based on convolutional neural networks (CNN). The results of this study can be a force for developing an extension model that can generate a set of sentences covering the same subjects, and furthermore, we can implement an artificial intelligence model that automatically creates technical literature.

• The Journal of the Acoustical Society of Korea
• /
• v.39 no.5
• /
• pp.497-504
• /
• 2020

Constructing speaker embeddings in speaker verification is an important issue. In general, a self-attention mechanism has been applied for speaker embedding encoding. Previous studies focused on training the self-attention in a high-level layer, such as the last pooling layer. In this case, the effect of low-level layers is not well represented in the speaker embedding encoding. In this study, we propose Masked Cross Self-Attentive Encoding (MCSAE) using ResNet. It focuses on training the features of both high-level and low-level layers. Based on multi-layer aggregation, the output features of each residual layer are used for the MCSAE. In the MCSAE, the interdependence of each input features is trained by cross self-attention module. A random masking regularization module is also applied to prevent overfitting problem. The MCSAE enhances the weight of frames representing the speaker information. Then, the output features are concatenated and encoded in the speaker embedding. Therefore, a more informative speaker embedding is encoded by using the MCSAE. The experimental results showed an equal error rate of 2.63 % using the VoxCeleb1 evaluation dataset. It improved performance compared with the previous self-attentive encoding and state-of-the-art methods.

• Journal of the Korea Society of Computer and Information
• /
• v.27 no.3
• /
• pp.13-23
• /
• 2022

In this study we leveraged a probing method to evaluate how a pre-trained L2 LSTM language model represents sentences with relative and coordinate clauses. The probing experiment employed adapted models based on the pre-trained L2 language models to trace the syntactic properties of sentence embedding vector representations. The dataset for probing was automatically generated using several templates related to different sentence structures. To classify the syntactic properties of sentences for each probing task, we measured the adaptation effects of the language models using syntactic priming. We performed linear mixed-effects model analyses to analyze the relation between adaptation effects in a complex statistical manner and reveal how the L2 language models represent syntactic features for English sentences. When the L2 language models were compared with the baseline L1 Gulordava language models, the analogous results were found for each probing task. In addition, it was confirmed that the L2 language models contain syntactic features of relative and coordinate clauses hierarchically in the sentence embedding representations.

• Journal of Intelligence and Information Systems
• /
• v.29 no.1
• /
• pp.27-40
• /
• 2023

With the increase in the spread of smart devices and the impact of COVID-19, the consumption of media contents through smart devices has significantly increased. Along with this trend, the amount of media contents viewed through OTT platforms is increasing, that makes contents recommendations on these platforms more important. Previous contents-based recommendation researches have mostly utilized metadata that describes the characteristics of the contents, with a shortage of researches that utilize the contents' own descriptive metadata. In this paper, various text data including titles and synopses that describe the contents were used to recommend similar contents. KLUE-RoBERTa-large, a Korean language model with excellent performance, was used to train the model on the text data. A dataset of over 20,000 contents metadata including titles, synopses, composite genres, directors, actors, and hash tags information was used as training data. To enter the various text features into the language model, the features were concatenated using special tokens that indicate each feature. The test set was designed to promote the relative and objective nature of the model's similarity classification ability by using the three contents comparison method and applying multiple inspections to label the test set. Genres classification and hash tag classification prediction tasks were used to fine-tune the embeddings for the contents meta text data. As a result, the hash tag classification model showed an accuracy of over 90% based on the similarity test set, which was more than 9% better than the baseline language model. Through hash tag classification training, it was found that the language model's ability to classify similar contents was improved, which demonstrated the value of using a language model for the contents-based filtering.