Pytorch implementation for codes in Learning Modality-Specific Representations with Self-Supervised Multi-Task Learning for Multimodal Sentiment Analysis (AAAI2021). Please see our another repo MMSA for more details, which is a scalable framework for MSA.

1. Datasets and pre-trained berts

Download dataset features and pre-trained berts from the following links.

• Baidu Cloud Drive with code: ctgs
• Google Cloud Drive

For all features, you can use SHA-1 Hash Value to check the consistency.

MOSI/unaligned_50.pkl: 5da0b8440fc5a7c3a457859af27458beb993e088
MOSI/aligned_50.pkl: 5c62b896619a334a7104c8bef05d82b05272c71c
MOSEI/unaligned_50.pkl: db3e2cff4d706a88ee156981c2100975513d4610
MOSEI/aligned_50.pkl: ef49589349bc1c2bc252ccc0d4657a755c92a056
SIMS/unaligned_39.pkl: a00c73e92f66896403c09dbad63e242d5af756f8

Due to the size limitations, the MOSEI features and SIMS raw videos are available in Baidu Cloud Drive only. All dataset features are organized as:

{
    "train": {
        "raw_text": [],
        "audio": [],
        "vision": [],
        "id": [], # [video_id$_$clip_id, ..., ...]
        "text": [],
        "text_bert": [],
        "audio_lengths": [],
        "vision_lengths": [],
        "annotations": [],
        "classification_labels": [], # Negative(< 0), Neutral(0), Positive(> 0)
        "regression_labels": []
    },
    "valid": {***}, # same as the "train" 
    "test": {***}, # same as the "train"
}

For MOSI and MOSEI, the pre-extracted text features are from BERT, different from the original glove features in the CMU-Multimodal-SDK.

For SIMS, if you want to extract features from raw videos, you need to install Openface Toolkits first, and then refer our codes in the data/DataPre.py.

python data/DataPre.py --data_dir [path_to_Dataset] --language ** --openface2Path  [path_to_FeatureExtraction]

For bert models, you also can download Bert-Base, Chinese from Google-Bert. And then, convert tensorflow into pytorch using transformers-cli

1. Clone this repo and install requirements.

git clone https://github.com/thuiar/Self-MM
cd Self-MM
conda create --name self_mm python=3.7
source activate self_mm
pip install -r requirements.txt

1. Make some changes Modify the config/config_tune.py and config/config_regression.py to update dataset pathes.

2. Run codes

python run.py --modelName self_mm --datasetName mosi

Detailed results are shown in MMSA > results/result-stat.md.

Please cite our paper if you find our work useful for your research:

@inproceedings{yu2021le,
  title={Learning Modality-Specific Representations with Self-Supervised Multi-Task Learning for Multimodal Sentiment Analysis},
  author={Yu, Wenmeng and Xu, Hua and Ziqi, Yuan and Jiele, Wu},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  year={2021}
}