This paper introduces a hybrid dual-path architecture that bridges the gap between interpretable lexicon-based methods and opaque LLMs by integrating a contextual encoder (RoBERTa-MLP) with a graph-based encoder (GAT-eMFD) that captures structured moral knowledge. We further extend this into a multimodal framework, MOTIV, which incorporates textual, spatial, temporal, and behavioral data to contextualize moral expression. We evaluate our approach on three diverse corpora—MFTC (Twitter/X, 35,108 tweets), MFRC (Reddit, 16,123 comments), and the geotagged MOTIV dataset (1,483 Geo-tagged tweets)—unified under a consistent moral foundation schema. Our dual-path model sets a new state of the art, achieving a Macro F1-score of 0.69 on MFTC (a 3% improvement over the BERT baseline of 0.67) and 0.40 on MFRC. Ablation studies confirm that the performance gains arise from the fusion of contextual semantics and moral knowledge graphs. However, severe class imbalance in the datasets particularly affecting underrepresented foundations like Fairness and Purity remains a critical limitation, as reflected in 0.00 F1 scores for these classes. This work offers two key contributions: (1) demonstrating the effectiveness of combining deep semantics with structured moral reasoning, and (2) presenting a multimodal paradigm for analyzing moral discourse. These findings advance moral foundation prediction and highlight ongoing challenges in achieving fair and reliable moral inference in AI systems.
