MRER

Mitigating Alignment Bias in Multimodal Sentiment Analysis via Reliability-Aware Fusion and Evidence-Preserving Reconstruction.MRER, a modality reliability-aware evidencerecoverable framework for multimodal sentiment analysis. MR-PGF (Modality Reliability-aware Public Gated Fusion): Estimates sample-adaptive modality reliability from public representations and performs gated fusion with anti-collapse regularization, reducing alignment bias toward dominant modalities. ER-DCA (Evidence-Recoverable Decoding and Decision-Consistency Alignment): Constrains decomposed representations to remain reconstructive, semantically faithful, and prediction-consistent, improving the recoverability and usability of compressed evidence.

Project Structure

MRER/
├── train.py                  # Training entry
├── test.py                   # Evaluation script
├── run.py                    # Main runner
├── config/
│   └── config.json           # Runtime configuration
├── trains/
│   └── singleTask/
│       ├── MRER.py           # Trainer
│       └── model/
│           └── mrer.py       # Core MRER architecture
├── models/
│   └── cross_modal_ssm.py    # Cross-modal interaction module
├── dataset/                  # Processed MOSI/MOSEI datasets
├── utils/                    # Logging and evaluation utilities
└── requirements.txt

Datasets

• CMU-MOSI
• CMU-MOSEI

Place datasets in the ./dataset folder, or modify the dataset path in config/config.json.

Installation

1. Create a virtual environment (recommended):

python3 -m venv mrer_env
source mrer_env/bin/activate

2. Install dependencies:

pip install -r requirements.txt

Execution

Training

Set dataset_name='mosi' or dataset_name='mosei' in train.py, then run:

python train.py

The trained model will be saved in the ./pt directory.

Testing

Set the dataset name in test.py and the model path in run.py, then run:

python test.py

Configuration

Runtime parameters can be modified in ./config/config.json, including:

• Dataset paths
• Hyperparameters (learning rate, batch size, etc.)
• Module switches (MR-PGF, ER-DCA)

Logs and results are saved in ./log and ./result/normal directories.