DeepCU-IJCAI19

DeepCU: Integrating Both Common and Unique Latent Information for Multimodal Sentiment Analysis, IJCAI-19

Citation

When using this code, or the ideas of DeepCU, please cite the following paper

@INPROCEEDINGS{Verma0ZL19ijcai19,
 author = {Sunny Verma and Chen Wang and Liming Zhu and Wei Liu},
 title = DeepCU: Integrating both Common and Unique Latent Information for Multimodal Sentiment Analysis},
 booktitle = {Proceedings of the Twenty-Eighth International Joint Conference on Artificial Intelligence, {IJCAI} Macao},
 pages     = {3627--3634},
 year      = {2019},
 }

Dependencies / Notes

DeepCU is written in python3 with some code fragments copied from DCGAN implementation from Carpedm20.

• The code is developed with Python 3.6 and TensorFlow 1.12.0 (with GPU support) on Linux
• For reasons of my convenience, data_dir is required to be data_dir = ../../data -- errors might pop-up when other directories are used.
• The experiments (main.py) - loads pretrained model and executes test (i.e. prediction with our trained model). If you wish to train on your own data, please edit as deep_cu.train(FLAGS).

Future research (ideas)

• Better Fusion scheme for utilizing both common and unique latent information
• Utilize Sequence information for sentiment prediction
• Cross Data Generelaization Performance

Please contact either Sunny Verma or Wei Liu at [email protected] if you're interested to collaborate on this!