• Proceedings of the KSPS conference
• /
• 2005.11a
• /
• pp.103-106
• /
• 2005

In this paper, we propose a language modeling approach to improve the performance of a large vocabulary continuous speech recognition system. The proposed approach is based on the active learning framework that helps to select a text corpus from a plenty amount of text data required for language modeling. The perplexity is used as a measure for the corpus selection in the active learning. From the recognition experiments on the task of continuous Korean speech, the speech recognition system employing the language model by the proposed language modeling approach reduces the word error rate by about 6.6 % with less computational complexity than that using a language model constructed with randomly selected texts.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.16 no.7
• /
• pp.2407-2424
• /
• 2022

Recently, Transformer has made great progress in video retrieval tasks due to its high representation capability. For the structure of a Transformer, the cascaded self-attention modules are capable of capturing long-distance feature dependencies. However, the local feature details are likely to have deteriorated. In addition, increasing the depth of the structure is likely to produce learning bias in the learned features. In this paper, an improved Transformer structure named TransDCS (Transformer with Dynamic Convolution and Shortcut) is proposed. A Multi-head Conv-Self-Attention module is introduced to model the local dependencies and improve the efficiency of local features extraction. Meanwhile, the augmented shortcuts module based on a dual identity matrix is applied to enhance the conduction of input features, and mitigate the learning bias. The proposed model is tested on MSRVTT, LSMDC and Activity-Net benchmarks, and it surpasses all previous solutions for the video-text retrieval task. For example, on the LSMDC benchmark, a gain of about 2.3% MdR and 6.1% MnR is obtained over recently proposed multimodal-based methods.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.10 no.9
• /
• pp.4568-4584
• /
• 2016

Steganography based on text generation has become a hot research topic in recent years. However, current text-generation methods which generate texts of normal style have either semantic or syntactic flaws. Note that texts of special genre, such as poem, have much simpler language model, less grammar rules, and lower demand for naturalness. Motivated by this observation, in this paper, we propose a text steganography that utilizes Markov chain model to generate Ci-poetry, a classic Chinese poem style. Since all Ci poems have fixed tone patterns, the generation process is to select proper words based on a chosen tone pattern. Markov chain model can obtain a state transfer matrix which simulates the language model of Ci-poetry by learning from a given corpus. To begin with an initial word, we can hide secret message when we use the state transfer matrix to choose a next word, and iterating until the end of the whole Ci poem. Extensive experiments are conducted and both machine and human evaluation results show that our method can generate Ci-poetry with higher naturalness than former researches and achieve competitive embedding rate.

• Smart Media Journal
• /
• v.12 no.7
• /
• pp.76-83
• /
• 2023

This study proposes a method to assist webtoon artists in the process of webtoon creation by utilizing a pretrained Text-to-Image model to generate webtoon images from text. The proposed approach involves fine-tuning a pretrained Stable Diffusion model using a webtoon dataset transformed into the desired webtoon style. The fine-tuning process, using LoRA technique, completes in a quick training time of approximately 4.5 hours with 30,000 steps. The generated images exhibit the representation of shapes and backgrounds based on the input text, resulting in the creation of webtoon-like images. Furthermore, the quantitative evaluation using the Inception score shows that the proposed method outperforms DCGAN-based Text-to-Image models. If webtoon artists adopt the proposed Text-to-Image model for webtoon creation, it is expected to significantly reduce the time required for the creative process.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.6 no.11
• /
• pp.2866-2879
• /
• 2012

Among the various security threats in online games, the use of game bots is the most serious problem. Previous studies on game bot detection have proposed many methods to find out discriminable behaviors of bots from humans based on the fact that a bot's playing pattern is different from that of a human. In this paper, we look at the chatting data that reflects gamers' communication patterns and propose a communication pattern analysis framework for online game bot detection. In massive multi-user online role playing games (MMORPGs), game bots use chatting message in a different way from normal users. We derive four features; a network feature, a descriptive feature, a diversity feature and a text feature. To measure the diversity of communication patterns, we propose lightly summarized indices, which are computationally inexpensive and intuitive. For text features, we derive lexical, syntactic and semantic features from chatting contents using text mining techniques. To build the learning model for game bot detection, we test and compare three classification models: the random forest, logistic regression and lazy learning. We apply the proposed framework to AION operated by NCsoft, a leading online game company in Korea. As a result of our experiments, we found that the random forest outperforms the logistic regression and lazy learning. The model that employs the entire feature sets gives the highest performance with a precision value of 0.893 and a recall value of 0.965.

• Journal of the Korea Society of Computer and Information
• /
• v.29 no.4
• /
• pp.23-30
• /
• 2024

In this paper, we propose a deep learning-based personalized senior care service application. The proposed application uses Speech to Text technology to convert the user's speech into text and uses it as input to Autogen, an interactive multi-agent large-scale language model developed by Microsoft, for user convenience. Autogen uses data from previous conversations between the senior and ChatBot to understand the other user's intent and respond to the response, and then uses a back-end agent to create a wish list, a shared calendar, and a greeting message with the other user's voice through a deep learning model for voice cloning. Additionally, the application can perform home IoT services with SKT's AI speaker (NUGU). The proposed application is expected to contribute to future AI-based senior care technology.

• 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.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.16 no.10
• /
• pp.3211-3229
• /
• 2022

To solve the problems existing in the process of Weibo disaster rumor recognition, such as lack of corpus, poor text standardization, difficult to learn semantic information, and simple semantic features of disaster rumor text, this paper takes Sina Weibo as the data source, constructs a dataset for Weibo disaster rumor recognition, and proposes a deep learning model BERT_AT_Stacked LSTM for Weibo disaster rumor recognition. First, add adversarial disturbance to the embedding vector of each word to generate adversarial samples to enhance the features of rumor text, and carry out adversarial training to solve the problem that the text features of disaster rumors are relatively single. Second, the BERT part obtains the word-level semantic information of each Weibo text and generates a hidden vector containing sentence-level feature information. Finally, the hidden complex semantic information of poorly-regulated Weibo texts is learned using a Stacked Long Short-Term Memory (Stacked LSTM) structure. The experimental results show that, compared with other comparative models, the model in this paper has more advantages in recognizing disaster rumors on Weibo, with an F1_Socre of 97.48%, and has been tested on an open general domain dataset, with an F1_Score of 94.59%, indicating that the model has better generalization.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.24 no.5
• /
• pp.670-673
• /
• 2020

Recent advances in artificial intelligence have allowed for easier sentiment analysis (e.g. positive or negative forecast) of documents such as a finance reports. In this paper, we investigate a method to apply text mining techniques to extract in the financial report using deep learning, and propose an accounting model for the effects of sentiment values in financial information. For sentiment analysis with keyword detection in the financial report, we suggest the input layer with extracted keywords, hidden layers by learned weights, and the output layer in terms of sentiment scores. Our approaches can help more effective strategy for potential investors as a professional guideline using sentiment values.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.10 no.6
• /
• pp.2527-2545
• /
• 2016

Measuring the similarity of given samples is a key problem of recognition, clustering, retrieval and related applications. A number of works, e.g. kernel method and metric learning, have been contributed to this problem. The challenge of similarity learning is to find a similarity robust to intra-class variance and simultaneously selective to inter-class characteristic. We observed that, the similarity measure can be improved if the data distribution and hidden semantic information are exploited in a more sophisticated way. In this paper, we propose a similarity learning approach for retrieval and recognition. The approach, termed as LDA-FEK, derives free energy kernel (FEK) from Latent Dirichlet Allocation (LDA). First, it trains LDA and constructs kernel using the parameters and variables of the trained model. Then, the unknown kernel parameters are learned by a discriminative learning approach. The main contributions of the proposed method are twofold: (1) the method is computationally efficient and scalable since the parameters in kernel are determined in a staged way; (2) the method exploits data distribution and semantic level hidden information by means of LDA. To evaluate the performance of LDA-FEK, we apply it for image retrieval over two data sets and for text categorization on four popular data sets. The results show the competitive performance of our method.