Hate speech, defined as derogatory statements directed against individuals or groups because of their immutable characteristics, has far-reaching negative consequences. The effects of hate speech conveyed in text (i.e., written hate speech) have been studied extensively. However, the role of other modalities, particularly in spoken form (i.e., spoken hate speech), remains underexplored. This is due to the lack of annotated multimodal databases and the ambiguity that arises from varying definitions of hate speech. The present work introduces the Warsaw Multimodal Hate Speech Database (WMHS), a new resource designed to fill these gaps. The database contains over 9 hours of manually annotated Polish hate speech, primarily targeting six social groups, collected from YouTube, Facebook, and the Polish platform BanBye. We present initial validation results of two pilot studies (N = 20), demonstrating high levels of offensiveness and emotional charge. We also describe results for automatic detection of hate speech, with accuracies up to 95.2% using linguistic features and 86.5% using acoustic features. The fusion model surpassed both unimodal baselines, reaching 96.5% accuracy and indicating that speech provides information complementary to text. We discuss how the WMHS can support a broad range of future work, including training machine learning models and psychological research on reactions to hate speech.
This paper proposes a model architecture named Reinforcement Adaptation with Dynamic Alignment and Reversal (R-DARe), which integrates reinforcement learning with transfer learning to address data imbalance and domain shift in cross-subject emotion recognition based on electroencephalogram (EEG) signals. The framework is evaluated under a calibration-assisted semi-supervised transductive cross-subject adaptation protocol, in which a small labeled target-domain calibration subset provides limited supervision, while emotion-unlabeled target-test EEG inputs are accessible during adaptation. The model formulates source-domain sample selection as a reinforcement learning problem optimized via REINFORCE. By constructing a composite reward function, balanced accuracy (bACC) and Macro-F1 are used as performance feedback, and the global maximum mean discrepancy (MMD) and KL divergence are incorporated as distributional constraints, thereby dynamically guiding the agent to learn and adjust the sampling strategy. This incentivizes the model to proactively adapt to target domain requirements and achieve dynamic alignment with the target domain, mitigating negative transfer effects caused by majority class dominance. R-DARe employs Parallel Multi-Scale Spectral Blocks (PMSB) to capture inter-band correlations, and integrates an Angular-Margin Classification Head (AMCH) with MMD techniques to further enhance inter-class separability and domain invariance. Experiments on the Cross-Subject EEG Arousal Dataset (CEAD) and the SEED dataset demonstrate that R-DARe achieves the highest mean performance compared with other data- imbalance handling methods. These findings suggest the potential of R-DARe for addressing data imbalance in cross-subject EEG emotion recognition.
Despite the challenges posed by electroencephalogram (EEG) variability, recent advances have been made in personalized affective brain-computer interfaces (aBCIs). Emotion-annotated EEG repositories are a critical driving force behind these advances. However, data-label ambiguity may arise from the lingering and temporal accumulative effects of emotion elicitation during data collection. Such unsolved ambiguity may impose a substantial burden on emotion modeling, particularly for datasets spanning multiple session across days. This study introduced an upstream data driven screening pipeline for filtering EEG trials based on their emotional validity prior to emotion modeling. A new metric, the trial-based emotional validity index (tEVI), was proposed to quantify the disparity between each emotion-elicitation EEG trial and its corresponding baseline. Using add-day-in (ADI) validation to an eight-day dataset collected over two months, the results showed that retaining EEG trials with less data-label ambiguity improved the generalizability of both the feature space and predictive models for cross-day emotion classification. With robust principal component analysis, the most stringent tEVI threshold increased the average ADI balanced accuracy across seven ADI configurations by 7.6% for valence and 4.6% for arousal. This study provides the first systematic validation of an emotional validity-driven upstream preprocessing framework using a longitudinal EEG dataset and highlights its potential for future integration with advanced affective-computing methods to facilitate robust emotion classification in real-world aBCI settings.
Personalised music-based interventions offer a powerful means of supporting motor rehabilitation by dynamically tailoring auditory stimuli to provide external timekeeping cues, modulate affective states, and stabilise gait patterns. Brain- Computer Interfaces (BCIs) thus hold promise for detecting EEG activity associated with adaptative and non-adaptive gait patterns, enabling dynamic modifications of rhythmic structures and the personalisation of such interventions across individuals. However, inter-subject variability in EEG signals, further compounded by movement-induced artefacts and motor planning differences, hinders the generalisability of BCIs and results in lengthy calibration processes. We propose Individual Tangent Space Alignment (ITSA), a novel pre-alignment strategy incorporating subject-specific recentering, distribution matching, and supervised rotational alignment to enhance cross-subject generalisation. Our hybrid architecture fuses Regularised Common Spatial Patterns (RCSP) with Riemannian geometry in parallel and sequential configurations, improving class separability while maintaining the geometric structure of covariance matrices for robust statistical computation. Using leave-one-subject-out cross-validation, ‘ITSA’ demonstrates significant performance improvements across subjects and conditions. The parallel fusion approach shows the greatest enhancement over its sequential counterpart, with robust performance maintained across varying data conditions and electrode configurations. The accompany code scripts can be found: https://github.com/NLai-CDT/ITSA.git
Accurate emotion recognition from electroencephalogram (EEG) signals is a key challenge in affective computing, due to EEG's non-stationary nature and the complex interactions among frequency, temporal, and spatial domains. Existing methods often separate these domains, failing to capture their inherent synergy during emotional processing. To address this, we propose the Frequency-aware Spatio-Temporal Hypergraph Network (FSTHG)-a novel end-to-end framework that unifies frequency, time, and space into a single hypergraph representation. In FSTHG, each node encodes EEG features across channels, time, and frequency bands, while hyperedges model high-order dependencies among them. Our key innovations include: (1) A wavelet-based signal decomposition module that extracts multi-scale frequency components, distinguishing between sustained rhythms and transient bursts; (2) A dual-branch architecture that separately processes high- and low-frequency components and enhances cross-frequency interactions through intra-branch fusion; (3) A novel Dynamic Fourier Hypergraph Convolution (DFHConv) module that operates in the frequency domain using node-invariant dynamic filters, enabling efficient and adaptive modeling of complex spatio-temporal relations. Comprehensive experiments on three benchmark EEG datasets (DEAP, DREAMER, and SEED) show that FSTHG achieves competitive and consistently strong results in accuracy, stability, and computational efficiency among the compared methods. Furthermore, visualization analyses demonstrate that the learned features align with known neurophysiological patterns, validating the model's interpretability and its ability to capture frequency-time-space synergy in emotional brain responses.
