Multimodal sentiment analysis and emotion recognition require modeling heterogeneous interactions among text, acoustic and visual streams. Large language models (LLMs) excel at language understanding, yet adapting them to affective computing is difficult: full‑model fine‑tuning is computationally prohibitive and many lightweight adapters dilute rich textual cues during cross‑modal fusion. To overcome these issues we introduce the Multi‑View Text‑Guided Multimodal Fusion Adapter (MVFA), a parameter‑efficient framework that augments frozen LLMs with strong multimodal reasoning. MVFA first generates complementary text views via max‑pooling, mean‑pooling and attention‑pooling; these views guide cross‑modal interactions with audio and visual features. The fused multimodal representation is then compressed into a compact set of learnable pseudo‑tokens by an Enhanced Q‑Former Fusion Module. Using ChatGLM3‑6B‑base as the primary backbone and additionally testing LLaMA2‑7B and Qwen3‑8B demonstrates the method’s portability across frozen LLMs. Experiments on CH‑SIMS V2.0, MELD and CHERMA show state‑of‑the‑art results: 84.62% Acc2 and 84.59% F1 on CH‑SIMS V2.0, 67.36% Acc and 66.03% WF1 on MELD, and 74.66% Acc on CHERMA, while updating only a tiny fraction of parameters. These results establish multi‑view text‑guided fusion as an effective and scalable paradigm for parameter‑efficient multimodal LLM adaptation in affective computing. The code is publicly available at https://github.com/Overwhelm1208/MVFA.
Review