Existing multimodal sentiment analysis (MSA) methods rely on representation learning and cross-modal fusion. However, many follow a correlation-driven paradigm that may encode lexical shortcuts, visual biases, and non-causal correlations into predictions. To address this, we propose a Consistency-Conflict Fusion with Sample-Adaptive Causal Evidence (CCF-SACE) framework, revisiting MSA from a representation-level causal modeling perspective. First, encoders extract textual and visual features, and an adaptor projects them into a shared space. The framework then performs sample-adaptive cross-modal causal evidence selection, where textual causal evidence is selected under visual context and visual causal evidence under textual context. The remaining contexts are modeled as non-causal proxies, allowing the model to separate emotion-supportive evidence from shortcut-related information. CCF-SACE further introduces a covariance-based consistency-conflict-aware fusion module that constructs an evidence-level mediator by modeling cross-modal consistency and discrepancy rather than directly fusing all selected cues. Since the residual proxy stream may still carry label-correlated shortcut information, we introduce non-causal proxy suppression and counterfactual sensitivity objectives to weaken its standalone predictability and enlarge the discrepancy between evidence-mediated and proxy-only predictions. Experiments on five benchmarks show consistent improvements over state-of-the-art methods, especially on fine-grained affective recognition.

The dimensional approach can represent more fine-grained emotional information than discrete affective states. In this paper, a pretrained language model (PLM) with a joint learning strategy is proposed for the SIGHAN-2024 shared task on Chinese dimensional aspect-based sentiment analysis (dimABSA), which requires submitted models to provide fine-grained multi-dimensional (Valance and Arousal) intensity predictions for given aspects of a review. The proposed model consists of three parts: an input layer that concatenates both given aspect terms and input sentences; a Chinese PLM encoder that generates aspect-specific review representation; and separate linear predictors that jointly predict Valence and Arousal sentiment intensities. Moreover, we merge simplified and traditional Chinese training data for data augmentation. Our systems ranked 2nd place out of 5 participants in subtask 1-intensity prediction. The code is publicly available at https://github.com/WZH5127/2024_subtask1_intensity_prediction.