Article
Computer Science, Artificial Intelligence
Lea Canales, Walter Daelemans, Ester Boldrini, Patricio Martinez-Barco
Summary: The exponential growth of subjective information on Web 2.0 has led to increased interest in developing methods to extract emotion data from these new sources. This study proposes EmoLabel, a semi-automatic methodology based on pre-annotation, and evaluates the impact of this method on manual emotion annotation in terms of agreement and annotation time. The results show that pre-annotation processes provide benefits in terms of annotation time without compromising annotator performance, particularly for inaccurate annotators.
IEEE TRANSACTIONS ON AFFECTIVE COMPUTING
(2022)
Article
Computer Science, Artificial Intelligence
Mirko Agarla, Simone Bianco, Luigi Celona, Paolo Napoletano, Alexey Petrovsky, Flavio Piccoli, Raimondo Schettini, Ivan Shanin
Summary: Thanks to deep learning techniques, performance in Speech Emotion Recognition(SER) on a single language has greatly increased in recent years. However, cross-lingual SER remains a challenge in real world applications due to the big gap among source and target domain distributions and the availability of unlabeled utterances. To address this, a Semi-Supervised Learning (SSL) method based on Transformer is proposed, which adapts to the new domain by exploiting pseudo-labeling strategy on the unlabeled utterances, achieving robustness across five languages and significant improvement in unweighted accuracy compared to state-of-the-art methods.
EXPERT SYSTEMS WITH APPLICATIONS
(2024)
Article
Computer Science, Artificial Intelligence
Yunze Gao, Yingying Chen, Jinqiao Wang, Hanqing Lu
Summary: In this study, a novel semi-supervised method for scene text recognition is proposed, which optimizes the quality of generated strings by designing global metrics and using reinforcement learning techniques. The approach effectively utilizes unlabeled data and performs well on multiple benchmark datasets by evaluating the quality of generated strings through embedding rewards and edit rewards.
IEEE TRANSACTIONS ON IMAGE PROCESSING
(2021)
Article
Computer Science, Information Systems
Tianhui Sha, Yong Peng
Summary: Due to its temporal resolution, portability and low cost, electroencephalogram (EEG) signals have become increasingly popular in emotion recognition. However, the widely used least square regression (LSR) method has limitations in retaining discriminative information and using discrete labels for multiple emotion classification. To address these issues, an orthogonal semi-supervised regression with adaptive label dragging model (OSRLD) is proposed. Experimental results using the SEED-IV dataset show that OSRLD achieves the highest classification accuracy and can automatically identify primary EEG frequency bands and brain areas.
JOURNAL OF KING SAUD UNIVERSITY-COMPUTER AND INFORMATION SCIENCES
(2023)
Review
Computer Science, Artificial Intelligence
Jose Marcio Duarte, Lilian Berton
Summary: A large amount of data is generated daily, leading to challenges in handling big data. One of the challenges is in text mining, particularly text classification. Semi-supervised learning (SSL), which utilizes labeled and unlabeled data, has become increasingly important in this field. This paper aims to fill the gap by providing an up-to-date review of SSL for text classification, analyzing the application domain, datasets, languages, text representations, machine learning algorithms, evaluation metrics, and future trends.
ARTIFICIAL INTELLIGENCE REVIEW
(2023)
Article
Social Sciences, Interdisciplinary
Ziyuan Chen, Shuzhe Duan, Yong Peng
Summary: This paper proposes a new model RSRRW for emotion recognition using EEG data. The model improves the accuracy by adding probability weights, extending the epsilon-dragging method, and adaptively measuring the contribution of sample features.
Article
Computer Science, Artificial Intelligence
Sheng Zhang, Min Chen, Jincai Chen, Yuan-Fang Li, Yiling Wu, Minglei Li, Chuanbo Zhu
Summary: Speech emotion recognition faces challenges due to limited large, high-quality labeled datasets and noisy pseudo-labels. This study proposes a new architecture that combines cross-modal knowledge transfer from visual to audio modality in a semi-supervised learning method to alleviate these issues. Experiments on CH-SIMS and IEMOCAP datasets demonstrate that the proposed method outperforms state-of-the-art results by effectively utilizing additional unlabeled audio-visual data.
KNOWLEDGE-BASED SYSTEMS
(2021)
Article
Engineering, Electrical & Electronic
Xing Li, Fangyao Shen, Yong Peng, Wanzeng Kong, Bao-Liang Lu
Summary: Recently, there has been increasing interest in research on emotion recognition based on electroencephalogram (EEG). The properties of weak, non-stationary, multi-rhythm and multi-channel EEG data easily lead to different contributions of extracted EEG samples and features in recognizing emotional states. However, existing studies have either neglected the importance of both samples and features or only considered one of them. In this study, a new model called sJSFE (semi-supervised Joint Sample and Feature importance Evaluation) is proposed to quantitatively measure the importance of samples and features using self-paced learning and feature self-weighting, respectively. Experimental results on the SEED-IV dataset demonstrate that mining both sample and feature importance greatly improves the performance of emotion recognition. The average accuracy obtained by sJSFE across the three cross-session recognition tasks is 82.45%, which is significantly higher than the results of traditional models. Furthermore, the feature importance vector reveals that the Gamma frequency band contributes the most, and specific brain regions are correlated with emotion recognition.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS II-EXPRESS BRIEFS
(2022)
Article
Computer Science, Information Systems
Jiayi Zhang, Xingzhi Wang, Dong Zhang, Dah-Jye Lee
Summary: The performance of learning-based group emotion recognition methods relies on labeled samples. However, labeling group emotion images manually is costly, leading to small datasets that limit the performance. In this study, a semi-supervised framework based on contrastive learning is proposed to improve GER performance by utilizing both labeled and unlabeled images. The proposed method pretrains the backbone using contrastive learning on unlabeled images and fine-tunes the network using labeled images. Pseudo-labels are assigned to unlabeled images by the fine-tuned network for further training, while a Weight Cross-Entropy Loss is introduced to reduce the impact of unreliable pseudo-labels.
Article
Engineering, Electrical & Electronic
Yong Peng, Wanzeng Kong, Feiwei Qin, Feiping Nie, Jinglong Fang, Bao-Liang Lu, Andrzej Cichocki
Summary: The proposed self-weighted semi-supervised classification model successfully addressed the issues in EEG-based cross-session emotion recognition and activation patterns mining. Experimental results showed that the Gamma frequency band is the most crucial for emotion recognition, and specific EEG channels play key roles in cross-session emotion recognition.
IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT
(2021)
Article
Neurosciences
Yufang Dan, Jianwen Tao, Jianjing Fu, Di Zhou
Summary: The paper introduces a Possibilistic Clustering-Promoting semi-supervised learning method, which improves the reliability and robustness of EEG-based emotion recognition by utilizing local weighted mean and a regularization term about fuzzy entropy.
FRONTIERS IN NEUROSCIENCE
(2021)
Article
Computer Science, Artificial Intelligence
N. S. B. D. Kavitha, Prasad P. V. G. D. Reddy, K. Venkata Rao
Summary: Microblogging has become a popular social tool among internet users for sentiment analysis and opinion mining. This article focuses on Twitter and presents a unique approach for sentiment classification and mining of political tweets, achieving high accuracy and effectiveness.
JOURNAL OF EXPERIMENTAL & THEORETICAL ARTIFICIAL INTELLIGENCE
(2023)
Article
Computer Science, Information Systems
Fengzhe Jin, Yong Peng, Feiwei Qin, Junhua Li, Wanzeng Kong
Summary: In this paper, a Graph Adaptive Semi-supervised Discriminative Subspace Learning (GASDSL) model is proposed for EEG-based emotion recognition. GASDSL aims to explore a discriminative subspace that improves emotion recognition accuracy. Comparative studies show that GASDSL achieves satisfactory results compared to other semi-supervised learning models.
JOURNAL OF KING SAUD UNIVERSITY-COMPUTER AND INFORMATION SCIENCES
(2023)
Article
Chemistry, Analytical
Whanhee Cho, Yongsuk Choi
Summary: This paper introduces a semi-supervised learning approach for text classification called LMGAN, based on generative adversarial networks (GAN). LMGAN utilizes BERT and GAN-BERT, multiple generators, and hidden layer outputs to enrich the distribution of fake data, addressing the limitation of early GAN-based methods in generating high-quality fake data. Experimental results demonstrate that LMGAN achieves better performance even with limited amounts of labeled data.
Article
Computer Science, Artificial Intelligence
Guangyi Zhang, Vandad Davoodnia, Ali Etemad
Summary: PAIRWISE ALIGNMENT OF REPRESENTATIONS FOR SEMI-SUPERVISED ELECTROENCEPHALOGRAM (EEG) LEARNING (PARSE) is a semi-supervised architecture for emotion recognition in EEG data. It reduces distribution mismatch between unlabeled and labeled data through pairwise representation alignment. Experimental results demonstrate that it achieves good performance with varying amounts of labeled samples.
IEEE TRANSACTIONS ON AFFECTIVE COMPUTING
(2022)