• Proceedings of the Korean Society of Computer Information Conference
• /
• 2022.07a
• /
• pp.43-44
• /
• 2022

기존 자연어 처리 모델은 문맥 단위 단어 임베딩을 처리하지 못하는 한계점을 가지고 있는 한편 최근 BERT 기반 사전학습 모델들은 문장 단위 임베딩이 가능하고 사전학습을 통해 학습 효율이 비약적으로 개선되었다는 특징이 있다. 본 논문에서는 사전학습 언어 모델들을 이용하여 음식점, 배달전문점 등 음식 업종에서 발생한 고객 발화 의도를 분류하고 모델별 성능을 비교하여 최적의 모델을 제안하고자 한다. 연구결과, 사전학습 모델의 한국어 코퍼스와 Vocab 사이즈가 클수록 고객의 발화 의도를 잘 예측하였다. 한편, 본 연구에서 발화자의 의도를 크게 문의와 요청으로 구분하여 진행하였는데, 문의와 요청의 큰 차이점인 '물음표'를 제거한 후 성능을 비교해본 결과, 물음표가 존재할 때 발화자 의도 예측에 좋은 성능을 보였다. 이를 통해 음식 업종에서 발화자의 의도를 예측하는 시스템을 개발하고 챗봇 시스템 등에 활용한다면, 발화자의 의도에 적합한 서비스를 정확하게 적시에 제공할 수 있을 것으로 기대한다.

• ETRI Journal
• /
• v.46 no.2
• /
• pp.277-289
• /
• 2024

In this paper, we describe a neural network-based application that recommends multiple items using dialog context input and simultaneously outputs a response sentence. Further, we describe a multi-item recommendation by specifying it as a set of clothing recommendations. For this, a multimodal fusion approach that can process both cloth-related text and images is required. We also examine achieving the requirements of downstream models using a pretrained language model. Moreover, we propose a gate-based multimodal fusion and multiprompt learning based on a pretrained language model. Specifically, we propose an automatic evaluation technique to solve the one-to-many mapping problem of multi-item recommendations. A fashion-domain multimodal dataset based on Koreans is constructed and tested. Various experimental environment settings are verified using an automatic evaluation method. The results show that our proposed method can be used to obtain confidence scores for multi-item recommendation results, which is different from traditional accuracy evaluation.

• Tuberculosis and Respiratory Diseases
• /
• v.86 no.3
• /
• pp.226-233
• /
• 2023

Background: Inactive or old, healed tuberculosis (TB) on chest radiograph (CR) is often found in high TB incidence countries, and to avoid unnecessary evaluation and medication, differentiation from active TB is important. This study develops a deep learning (DL) model to estimate activity in a single chest radiographic analysis. Methods: A total of 3,824 active TB CRs from 511 individuals and 2,277 inactive TB CRs from 558 individuals were retrospectively collected. A pretrained convolutional neural network was fine-tuned to classify active and inactive TB. The model was pretrained with 8,964 pneumonia and 8,525 normal cases from the National Institute of Health (NIH) dataset. During the pretraining phase, the DL model learns the following tasks: pneumonia vs. normal, pneumonia vs. active TB, and active TB vs. normal. The performance of the DL model was validated using three external datasets. Receiver operating characteristic analyses were performed to evaluate the diagnostic performance to determine active TB by DL model and radiologists. Sensitivities and specificities for determining active TB were evaluated for both the DL model and radiologists. Results: The performance of the DL model showed area under the curve (AUC) values of 0.980 in internal validation, and 0.815 and 0.887 in external validation. The AUC values for the DL model, thoracic radiologist, and general radiologist, evaluated using one of the external validation datasets, were 0.815, 0.871, and 0.811, respectively. Conclusion: This DL-based algorithm showed potential as an effective diagnostic tool to identify TB activity, and could be useful for the follow-up of patients with inactive TB in high TB burden countries.

• Journal of the Korea Convergence Society
• /
• v.13 no.2
• /
• pp.13-20
• /
• 2022

This study is to develop a named entity recognition model specialized in criminal investigation domains using deep learning techniques. Through this study, we propose a system that can contribute to analysis of crime for prevention and investigation using data analysis techniques in the future by automatically extracting and categorizing crime-related information from text-based data such as criminal judgments and investigation documents. For this study, the criminal investigation domain text was collected and the required entity name was newly defined from the perspective of criminal analysis. In addition, the proposed model applying KoELECTRA, a pre-trained language model that has recently shown high performance in natural language processing, shows performance of micro average(referred to as micro avg) F1-score 98% and macro average(referred to as macro avg) F1-score 95% in 9 main categories of crime domain NER experiment data, and micro avg F1-score 98% and macro avg F1-score 62% in 56 sub categories. The proposed model is analyzed from the perspective of future improvement and utilization.

• Journal of the Korea Convergence Society
• /
• v.12 no.7
• /
• pp.37-44
• /
• 2021

Quality estimation (QE) can evaluate the quality of machine translation output even for those who do not know the target language, and its high utilization highlights the need for QE. QE shared task is held every year at Conference on Machine Translation (WMT), and recently, researches applying Pretrained Language Model (PLM) are mainly being conducted. In this paper, we conduct a survey on the QE task and research trends, and we summarize the features of PLM. In addition, we used a multilingual BART model that has not yet been utilized and performed comparative analysis with the existing studies such as XLM, multilingual BERT, and XLM-RoBERTa. As a result of the experiment, we confirmed which PLM was most effective when applied to QE, and saw the possibility of applying the multilingual BART model to the QE task.

• 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.

• Proceedings of the Korean Institute of Information and Commucation Sciences Conference
• /
• 2022.10a
• /
• pp.81-83
• /
• 2022

Understanding research trends is necessary to select research topics and explore related works. Most researchers search representative keywords of interesting domains or technologies to understand research trends. However some conferences in artificial intelligence or data mining fields recently publish hundreds to thousands of papers for each year. It makes difficult for researchers to understand research trend of interesting domains. In our paper, we propose an automatic technology keyphrase extraction method to support researcher to understand research trend for each conference or journal. Keyphrase extraction that extracts important terms or phrases from a text, is a fundamental technology for a natural language processing such as summarization or searching, etc. Previous keyphrase extraction technologies based on pretrained language model extract keyphrases from long texts so performances are degraded in short texts like titles of papers. In this paper, we propose a techonolgy keyphrase extraction model that is robust in short text and considers the importance of the word.

• Journal of the Korea Convergence Society
• /
• v.12 no.5
• /
• pp.1-7
• /
• 2021

In the latest trend of machine translation research, the model is pretrained through a large mono lingual corpus and then finetuned with a parallel corpus. Although many studies tend to increase the amount of data used in the pretraining stage, it is hard to say that the amount of data must be increased to improve machine translation performance. In this study, through an experiment based on the mBART model using parallel corpus filtering, we propose that high quality data can yield better machine translation performance, even utilizing smaller amount of data. We propose that it is important to consider the quality of data rather than the amount of data, and it can be used as a guideline for building a training corpus.

• The Journal of the Institute of Internet, Broadcasting and Communication
• /
• v.24 no.1
• /
• pp.83-88
• /
• 2024

In this research, we propose a Semi-Supervised learning based railroad surface defect detection method. The Resnet50 model, pretrained on ImageNet, was employed for the training. Data without labels are randomly selected, and then labeled to train the ResNet50 model. The trained model is used to predict the results of the remaining unlabeled training data. The predicted values exceeding a certain threshold are selected, sorted in descending order, and added to the training data. Pseudo-labeling is performed based on the class with the highest probability during this process. An experiment was conducted to assess the overall class classification performance based on the initial number of labeled data. The results showed an accuracy of 98% at best with less than 10% labeled training data compared to the overall training data.

• Proceedings of the Korea Information Processing Society Conference
• /
• 2023.05a
• /
• pp.624-626
• /
• 2023

Credit scoring is a technique used by financial institutions to assess the creditworthiness of potential borrowers. This involves evaluating a borrower's credit history to predict the likelihood of defaulting on a loan. This paper presents an ensemble of two Transformer based models within a framework for discriminating the default risk of loan applications in the field of credit scoring. The first model is FinBERT, a pretrained NLP model to analyze sentiment of financial text. The second model is FT-Transformer, a simple adaptation of the Transformer architecture for the tabular domain. Both models are trained on the same underlying data set, with the only difference being the representation of the data. This multi-modal approach allows us to leverage the unique capabilities of each model and potentially uncover insights that may not be apparent when using a single model alone. We compare our model with two famous ensemble-based models, Random Forest and Extreme Gradient Boosting.