SemMSA checkpoints

Checkpoints of SemMSA: Latent Semantic-Aided Robust Multimodal Sentiment Analysis with Incomplete Data (paper, code, project page, data).

SemMSA constructs sentiment-relevant latent semantics with a frozen LLM (Qwen3-1.7B) through Cross-modal Semantic Refinement (CSR) and aligns them with language, visual and acoustic representations through Cross-modal Spectral Alignment (CSA).

Released checkpoints

Checkpoints for CH-SIMS are released; checkpoints for MOSI and MOSEI will be added later.

<dataset>/<config>/best_<metric>_<seed>.pth

<config> is the training configuration configs/<dataset>/<config>.yaml of the code repository. Following the evaluation protocol of LNLN and P-RMF, one checkpoint is kept for each metric; configs/<dataset>/eval.yaml lists the checkpoint used for every reported metric.

Config	Files (seeds 1111, 1112, 1113)	Used for
sims/main	best_Mult_acc_2_*.pth, best_MAE_*.pth	Acc-2, Corr
sims/f1	best_Mult_acc_2_*.pth	F1
sims/reg	best_Mult_acc_3_*.pth, best_MAE_*.pth	Acc-3, MAE
sims/acc5	best_Mult_acc_5_*.pth	Acc-5

The files contain only the trainable weights: the frozen Qwen3-1.7B (and BERT, when it is frozen) are loaded from pretrained/.

Usage

git clone https://github.com/peacelwh/SemMSA.git && cd SemMSA
pip install -r requirements.txt && pip install -e .
bash scripts/download_pretrained.sh
bash scripts/download_data.sh sims
bash scripts/download_checkpoints.sh sims   # this repository -> checkpoints/
bash scripts/evaluate.sh sims 0             # robust evaluation under missing rates 0.0-0.9

Citation

@article{li2026semmsa,
  title   = {SemMSA: Latent Semantic-Aided Robust Multimodal Sentiment Analysis with Incomplete Data},
  author  = {Li, Wenhao and Wu, Zhibin and Xiao, Chong and Wang, Qiangchang},
  journal = {arXiv preprint arXiv:2609.30238},
  year    = {2026}
}