트랜스포머 기반 모델을 이용한 연구논문 멀티레이블 주제영역 분류

김성현,김영민 한국경영공학회 2024 한국경영공학회지 Vol.29 No.4

Purpose To develop MLP and transformer-based models for the multi-label topic classification of research papers using abstract text. Methods Abstracts from 119,600 papers in the Computer Science category of arXiv were collected to create a multi-label dataset with up to three categories out of a total of 15 possible categories. Performance was evaluated by developing a baseline MLP model along with transformer-based models: BERT, RoBERTa, and DistillBERT. Results The transformer models outperformed the traditional MLP model. The DistillBERT model achieved the highest micro F1-score of 0.749, while the BERT model recorded macro and weighted F1-scores of 0.655 and 0.733, respectively. The RoBERTa model excelled in the samples method with a score of 0.772. Conclusion This study enables researchers to quickly explore recent findings and effectively identify their research topics. Additionally, it is expected to significantly contribute to the efficient sharing of academic knowledge and the revitalization of the research community.

SYMMER: A Systematic Approach to Multiple Musical Emotion Recognition

Lee, Jae-Sung,Jo, Jin-Hyuk,Lee, Jae-Joon,Kim, Dae-Won Korean Institute of Intelligent Systems 2011 INTERNATIONAL JOURNAL of FUZZY LOGIC and INTELLIGE Vol.11 No.2

Music emotion recognition is currently one of the most attractive research areas in music information retrieval. In order to use emotion as clues when searching for a particular music, several music based emotion recognizing systems are fundamentally utilized. In order to maximize user satisfaction, the recognition accuracy is very important. In this paper, we develop a new music emotion recognition system, which employs a multilabel feature selector and multilabel classifier. The performance of the proposed system is demonstrated using novel musical emotion data.

SYMMER: A Systematic Approach to Multiple Musical Emotion Recognition

Jaesung Lee,Jin-Hyuk Jo,Jae-Joon Lee,Daewon Kim 한국지능시스템학회 2011 INTERNATIONAL JOURNAL of FUZZY LOGIC and INTELLIGE Vol.11 No.2

Music emotion recognition is currently one of the most attractive research areas in music information retrieval. In order to use emotion as clues when searching for a particular music, several music based emotion recognizing systems are fundamentally utilized. In order to maximize user satisfaction, the recognition accuracy is very important. In this paper, we develop a new music emotion recognition system, which employs a multilabel feature selector and multilabel classifier. The performance of the proposed system is demonstrated using novel musical emotion data.

SYMMER: A Systematic Approach to Multiple Musical Emotion Recognition

이재성,조진혁,이재준,김대원 한국지능시스템학회 2011 INTERNATIONAL JOURNAL of FUZZY LOGIC and INTELLIGE Vol.11 No.2

Music emotion recognition is currently one of the most attractive research areas in music information retrieval. In order to use emotion as clues when searching for a particular music, several music based emotion recognizing systems are fundamentally utilized. In order to maximize user satisfaction, the recognition accuracy is very important. In this paper, we develop a new music emotion recognition system, which employs a multilabel feature selector and multilabel classifier. The performance of the proposed system is demonstrated using novel musical emotion data.

HKIB-20000 & HKIB-40075: Hangul Benchmark Collections for Text Categorization Research

Kim, Jin-Suk,Choe, Ho-Seop,You, Beom-Jong,Seo, Jeong-Hyun,Lee, Suk-Hoon,Ra, Dong-Yul Korean Institute of Information Scientists and Eng 2009 Journal of Computing Science and Engineering Vol.3 No.3

The HKIB, or Hankookilbo, test collections are two archives of Korean newswire stories manually categorized with semi-hierarchical or hierarchical category taxonomies. The base newswire stories were made available by the Hankook Ilbo (The Korea Daily) for research purposes. At first, Chungnam National University and KISTI collaborated to manually tag 40,075 news stories with categories by semi-hierarchical and balanced three-level classification scheme, where each news story has only one level-3 category (single-labeling). We refer to this original data set as HKIB-40075 test collection. And then Yonsei University and KISTI collaborated to select 20,000 newswire stories from the HKIB-40075 test collection, to rearrange the classification scheme to be fully hierarchical but unbalanced, and to assign one or more categories to each news story (multi-labeling). We refer to this modified data set as HKIB-20000 test collection. We benchmark a k-NN categorization algorithm both on HKIB-20000 and on HKIB-40075, illustrating properties of the collections, providing baseline results for future studies, and suggesting new directions for further research on Korean text categorization problem.

HKIB-20000 & HKIB-40075: Hangul Benchmark Collections for Text Categorization Research

Jinsuk Kim,Ho-Seop Choe,Beom-Jong You,Jeong-Hyun Seo,Suk-Hoon Lee,Dong-Yul Ra 한국정보과학회 2009 Journal of Computing Science and Engineering Vol.3 No.3

The HKIB, or Hankookilbo, test collections are two archives of Korean newswire stories manually categorized with semi-hierarchical or hierarchical category taxonomies. The base newswire stories were made available by the Hankook Ilbo (The Korea Daily) for research purposes. At first, Chungnam National University and KISTI collaborated to manually tag 40,075 news stories with categories by semi-hierarchical and balanced three-level classification scheme, where each news story has only one level-3 category (single-labeling). We refer to this original data set as HKIB-40075 test collection. And then Yonsei University and KISTI collaborated to select 20,000 newswire stories from the HKIB-40075 test collection, to rearrange the classification scheme to be fully hierarchical but unbalanced, and to assign one or more categories to each news story (multi-labeling). We refer to this modified data set as HKIB-20000 test collection. We benchmark a k-NN categorization algorithm both on HKIB-20000 and on HKIB-40075, illustrating properties of the collections, providing baseline results for future studies, and suggesting new directions for further research on Korean text categorization problem.

A Neural Network-Based Approach to Multiple Wheat Disease Recognition

Igor V. Arinichev,Sergey V. Polyanskikh,Irina V. Arinicheva,Galina V. Volkova,Irina P. Matveeva 한국지능시스템학회 2022 INTERNATIONAL JOURNAL of FUZZY LOGIC and INTELLIGE Vol.22 No.1

In this paper, modern computer vision methods are proposed for detecting multiple diseases in wheat leaves. The authors demonstrate that modern neural network architectures are capable of qualitatively detecting and classifying diseases, such as yellow spots, yellow rust, and brown rust, even in cases in which multiple diseases are simultaneously present on the plant. For certain classes of diseases, the main multilabel metrics (accuracy, micro-/macro-precision, recall, and F1-score) range from 0.95 to 0.99. This indicates the possibility of recognizing several diseases on a leaf with an accuracy equal to that of an expert phytopathologist. The architecture of the neural network used in this case is lightweight, which makes it possible to use offline on mobile devices.

다중 레이블 인문 데이터의 효과적인 특징 선별을 위한 이주 개체 정제연산 기반 다중 개체군 유전알고리즘

박민우 ( Park Minwoo ),이재성 ( Lee Jaesung ) 중앙대학교 인공지능인문학연구소 2021 인공지능인문학연구 Vol.7 No.-

Multi-label feature selection is a preprocessing method that can be used to analyze, for example multi-label humanity data. In particular, a multi-population genetic algorithm is verified to exhibit a better performance for identifying an appropriate subset compared with existing genetic algorithms in that a variety of populations was preserved, and premature convergence was prevented. However, with this method, the inflow of closely related features to multi-labels is unlikely to search for the solution. This study proposes an effective multi-population genetic algorithm for multi-label feature selection. In the proposed method, a multi-population genetic algorithm with a refinement process in migrated individuals maintains a variety of populations, promotes the inflow of features closely related to multi-labels, and ultimately enhances the search performance. Experimental results indicate that the proposed method exhibit better performance than the compared multi-population algorithms.