Sublepidodendron is a common megafossil plant in the Late Devonian of China, but historically the generic delimitation based on leaf bases masked its true systematic position. A reinvestigation of S. songziense from the Late Devonian Hsiehchingssu Formation, Hubei, China, provides new insights into its internal anatomy and reproductive morphology. This arborescent lycopsid is characterized by small, vertically elongated leaf bases arranged in spirals, presence of false leaf scars, possibly bearing separate cones, and association with a stigmarian rhizomorph. The potential for organic connections of these detached organ genera has been noted for other Sublepidodendron species. The anatomy of S. songziense axes from two levels reveals that the thinner axis may bear an ectophloic siphonostele with a filamentous pith and an outer cortex. The thicker axis has a siphonostele with a branch gap, two‐zoned pith with secondary thickenings, multiseriate rays across secondary xylem, a thick periderm, and primary and secondary tracheid walls characterized by “Williamson's striations.” Similarities to synapomorphies of Diaphorodendraceae and Lepidodendraceae suggest that S. songziense bears a closer affinity to Lepidodendrales rather than Protolepidodendrales, as formerly thought. Widespread occurrence of Sublepidodendron implies that phylogenetically advanced arborescent lycopsids must have diverged by the Late Devonian.
23 December 2024 in ACM Transactions on Multimedia Computing, Communications, and Applications
Advancements in Multimodal Sentiment Analysis (MSA) have predominantly focused on leveraging the interdependence of text, acoustic, and visual modalities to enhance sentiment prediction. However, efficiently integrating these modalities remains a challenge. In response, we investigate the design of a practical approach to control the information flow more accurately and propose DWGCI. This multimodal model employs systematic cross-modal interactions and gating mechanisms to analyze sentiment. It ensures comprehensive feature extraction across modalities by utilizing BERT for text features, COVAREP for acoustic features, and CNNs for visual features. To improve the common feature expression of each modality and increase the dynamic interaction between modalities, we introduced a Text-Guided Multimodal Fusion Module (TGMFM). Another integral part of our model is a unique Dynamic Gating Module (DGM) strategically positioned to follow cross-modal interactions. This mechanism represents a significant innovation that can adapt to modality-specific sentiment cues with unprecedented precision dynamically. The model can make nuanced distinctions between modalities, which significantly enhances the accuracy of sentiment prediction. Our DWGCI has been subjected to rigorous testing on two widely used datasets. It demonstrated its superior performance, established a benchmark in MSA for gate fusion innovation, and highlighted its transformative potential.
Multimodal sentiment analysis (MSA) has become increasingly important for understanding human emotions, with applications in areas such as human-computer interaction, social media analysis, and emotion recognition. MSA leverages multimodal data, including text, audio, and visual inputs, to achieve better performance in emotion recognition. However, existing methods face challenges, particularly when dealing with missing modalities. While some approaches attempt to handle modality dropout, they often fail to effectively recover missing information or account for the complex interactions between different modalities. Moreover, many models treat modalities equally, not fully utilizing the unique strengths of each modality. To address these limitations, we propose the Modality-Uncertainty-aware Knowledge Distillation Framework (MUKDF). Specifically, we introduce a modality random missing strategy that enhances the model's adaptability to uncertain modality scenarios. To further improve performance, we incorporate a Dual-Branch Modality Knowledge Extractor (DMKE) that balances feature contributions across modalities and a multimodal masked transformer (MMT) designed to capture nuanced interactions between modalities. Additionally, we present a contrastive feature-level and align-based representation distillation mechanism (CFD&ARD), which strengthens the alignment between teacher and student representations, ensuring effective knowledge transfer and improved robustness in learning. Comprehensive experiments conducted on two benchmark datasets demonstrate that MUKDF outperforms several baseline models, achieving superior performance not only under complete modality conditions but also in the more challenging scenario with incomplete modalities. This highlights the effectiveness of our framework in handling the uncertainty and complexities inherent in multimodal sentiment analysis.