• Annual Conference on Human and Language Technology
• /
• 2019.10a
• /
• pp.37-40
• /
• 2019

한국어 형태소 분석은 입력된 문장 내의 어절들을 지니는 최소의 단위인 형태소로 분리하고 품사 부착하는 작업을 의미한다. 기존 한국어 형태소 분석 방법은 음절 기반 연구가 주를 이루고 이를 순차 태깅 문제로 보고 SVM, CRF혹은 Bi-LSTM-CRF 등을 이용하거나 특정 음절에서 형태소의 경계를 결정하는 전이 기반 모델을 통해 분석하는 모델 등이 연구되었다. 최근 자연어 처리 연구에서 대용량 코퍼스로부터 문맥을 고려한 BERT 등의 언어 모델을 활용한 연구가 각광받고 있다. 본 논문에서는 음절 단위가 아닌 BERT를 이용한 Sub-word 기반 형태소 분석 방법을 제안하고 기분석 사전을 통해 분석하는 과정을 거쳐 세종 한국어 형태소 분석 데이터 셋에서 형태소 단위 F1 : 95.22%, 어절 정확도 : 93.90%의 성능을 얻었다.

• Proceedings of the Korean Society of Computer Information Conference
• /
• 2022.01a
• /
• pp.7-8
• /
• 2022

본 연구에서는 한국어 SNS 대화에서 나타나는 문장들의 감성을 분석하고자 신경망 모델을 활용하여 시스템을 구축하였다. 현재 해외 SNS 감성분석에 대한 연구는 많이 진행된 상황이지만, 한국어 범용 대화에 대해 적절한 모델이 무엇인지는 연구가 부족한 실정이었다. 따라서 한국어 대화에 적합한 모델을 채택해 보다 정확한 감성분석을 수행하였다. 이를 위해 한국어 SNS 대화 데이터에 대해 신경망 모델을 적용하여, 82% 성공률로 기존 모델 72% 성공률보다 훨씬 더 우수한 성능을 보였다. 또한 본 연구의 결과는 악플 추적 등 실용적인 분야에도 기여할 수 있다고 사료된다.

• ETRI Journal
• /
• v.46 no.2
• /
• pp.205-217
• /
• 2024

We propose a bandwidth prediction approach based on deep learning. The approach is intended to accurately predict the bandwidth of various types of mobile networks. We first use a machine learning technique, namely, the gradient boosting algorithm, to recognize the connected mobile network. Second, we apply a handover detection algorithm based on network recognition to account for vertical handover that causes the bandwidth variance. Third, as the communication performance offered by 3G, 4G, and 5G networks varies, we suggest a bidirectional long short-term memory model with time distribution for bandwidth prediction per network. To increase the prediction accuracy, pretraining and fine-tuning are applied for each type of network. We use a dataset collected at University College Cork for network recognition, handover detection, and bandwidth prediction. The performance evaluation indicates that the handover detection algorithm achieves 88.5% accuracy, and the bandwidth prediction model achieves a high accuracy, with a root-mean-square error of only 2.12%.

• Annual Conference on Human and Language Technology
• /
• 2018.10a
• /
• pp.176-178
• /
• 2018

트위터는 최대 140자의 단문을 주고받는 소셜 네트워크 서비스이다. 트위터의 해시 태그는 주로 문장의 핵심 단어나 주요 토픽 등을 링크하게 되는데 본 논문에서는 이러한 정보를 이용하여 키워드 추출에 활용한다. 문장을 Character CNN, Bi-LSTM을 통해 문장 표현을 얻어내고 각 Span에서 이러한 문장 표현을 활용하여 Span 표현을 생성한다. Span 표현을 이용하여 각 Span에 대한 Score를 얻고 높은 점수의 Span을 이용하여 키워드를 추출한다.

• Annual Conference on Human and Language Technology
• /
• 2018.10a
• /
• pp.19-22
• /
• 2018

본 논문에서는 기존 Stack Pointer Network의 의존 파싱 모델을 확장한 Bi-Stack Pointer Network를 제안한다. Stack Pointer Network는 기존의 Pointer Network에 내부 stack을 만들어 전체 문장을 읽어 dependency tree를 구성한다. stack은 tree의 깊이 우선 탐색을 통해 선정되고 Pointer Network는 stack의 top 단어(head)의 자식(child)을 선택한다. 제안한 모델은 기존의 Stack Pointer Network가 지배소(head)정보로 의존소(child)를 예측하는 부분에 Biaffine attention을 통해 의존소(child)에서 지배소(head)를 예측하는 방향을 추가하여 양방향 예측이 가능하게 한 모델이다. 실험 결과, 제안 Bi-Stack Pointer Network모델은 UAS 91.53%, LAS 90.93%의 성능을 보여주어 기존 최고 성능을 개선시켰다.

• Journal of Information Processing Systems
• /
• v.14 no.1
• /
• pp.150-161
• /
• 2018

We propose a novel deep neural network model for detecting human activities in untrimmed videos. The process of human activity detection in a video involves two steps: a step to extract features that are effective in recognizing human activities in a long untrimmed video, followed by a step to detect human activities from those extracted features. To extract the rich features from video segments that could express unique patterns for each activity, we employ two different convolutional neural network models, C3D and I-ResNet. For detecting human activities from the sequence of extracted feature vectors, we use BLSTM, a bi-directional recurrent neural network model. By conducting experiments with ActivityNet 200, a large-scale benchmark dataset, we show the high performance of the proposed DeepAct model.

• The Journal of the Acoustical Society of Korea
• /
• v.40 no.5
• /
• pp.466-472
• /
• 2021

Speaker Change Detection (SCD) refers to finding the moment when the main speaker changes from one person to the next in a speech conversation. In speaker change detection, difficulties arise due to overlapping speakers, inaccuracy in the information labeling, and data imbalance. To solve these problems, TIMIT corpus widely used in speech recognition have been concatenated artificially to obtain a sufficient amount of training data, and the detection of changing speaker has performed after identifying overlapping speakers. In this paper, we propose an speaker change detection system that considers the speaker overlapping. We evaluated and verified the performance using various approaches. As a result, a detection system similar to the X-Vector structure was proposed to remove the speaker overlapping region, while the Bi-LSTM method was selected to model the speaker change system. The experimental results show a relative performance improvement of 4.6 % and 13.8 % respectively, compared to the baseline system. Additionally, we determined that a robust speaker change detection system can be built by conducting related studies based on the experimental results, taking into consideration text and speaker information.

• Journal of Broadcast Engineering
• /
• v.24 no.4
• /
• pp.553-563
• /
• 2019

As the number of videos uploaded on live streaming platforms rapidly increases, the demand for providing highlight videos is increasing to promote viewer experiences. In this paper, we present novel methods for predicting highlights using chat logs and audio data in videos. The proposed models employ bi-directional LSTMs to understand the contextual flow of a video. We also propose to use the features over various time-intervals to understand the mid-to-long term flows. The proposed Our methods are demonstrated on e-Sports and baseball videos collected from personal broadcasting platforms such as Twitch and Kakao TV. The results show that the information from multiple time-intervals is useful in predicting video highlights.

• KSII Transactions on Internet and Information Systems (TIIS)
• /
• v.15 no.7
• /
• pp.2304-2320
• /
• 2021

Intelligently detecting anomalies in health sensor data streams (e.g., Electrocardiogram, ECG) can improve the development of E-health industry. The physiological signals of patients are collected through sensors. Timely diagnosis and treatment save medical resources, promote physical health, and reduce complications. However, it is difficult to automatically classify the ECG data, as the features of ECGs are difficult to extract. And the volume of labeled ECG data is limited, which affects the classification performance. In this paper, we propose a Generative Adversarial Network (GAN)-based deep learning framework (called CAB) for heart arrhythmia classification. CAB focuses on improving the detection accuracy based on a small number of labeled samples. It is trained based on the class-imbalance ECG data. Augmenting ECG data by a GAN model eliminates the impact of data scarcity. After data augmentation, CAB classifies the ECG data by using a Bidirectional Long Short Term Memory Recurrent Neural Network (Bi-LSTM). Experiment results show a better performance of CAB compared with state-of-the-art methods. The overall classification accuracy of CAB is 99.71%. The F1-scores of classifying Normal beats (N), Supraventricular ectopic beats (S), Ventricular ectopic beats (V), Fusion beats (F) and Unclassifiable beats (Q) heartbeats are 99.86%, 97.66%, 99.05%, 98.57% and 99.88%, respectively. Unclassifiable beats (Q) heartbeats are 99.86%, 97.66%, 99.05%, 98.57% and 99.88%, respectively.

• Journal of the Korea Institute of Information and Communication Engineering
• /
• v.25 no.1
• /
• pp.1-7
• /
• 2021

Investor mood from social media is gaining increasing attention for leading a price movement in stock market. Based on the behavioral finance theory, this study argues that sentiment extracted from social media using big data technique can predict a real-time (short-run) price momentum in Chinese stock market. Collecting Sina Weibo posts that related to COVID-19 using keyword method, a daily influential weighted sentiment factors is extracted from the sizable raw data of over 2 millions of posts. We examine one supervised and 4 unsupervised sentiment analysis model, and use the best performed word-frequency and BiLSTM mdoel. The test result shows a similar movement between stock price change and sentiment factor. It indicates that public mood extracted from social media can in some extent represent the investors' sentiment and make a difference in stock market fluctuation when people are concentrating on a special events that can cause effect on the stock market.