TomBERT

Adapting BERT for Target-Oriented Multimodal Sentiment Classification

• Dataset and codes for our IJCAI 2019 paper "Adapting BERT for Target-Oriented Multimodal Sentiment Classification" https://www.ijcai.org/proceedings/2019/0751.pdf

Author

Jianfei YU

jfyu@njust.edu.cn

Mar 02, 2020

Target-oriented Multimodal Sentiment Classification (TMSC), PyTorch Implementations.

Requirement

• PyTorch 1.0.0
• Python 3.7

Download tweet images and set up image path

• Step 1: Download each tweet's associated image via this link (https://drive.google.com/file/d/1PpvvncnQkgDNeBMKVgG2zFYuRhbL873g/view)
• Step 2: Change the image path in line 553 and line 555 of the "run_multimodal_classifier.py" file
• Step 3: Download the pre-trained ResNet-152 via this link (https://download.pytorch.org/models/resnet152-b121ed2d.pth)
• Setp 4: Put the pre-trained ResNet-152 model under the folder named "resnet"

Code Usage

(Optional) Preprocessing

• This is optional, because I have provided the pre-processed data under the folder named "absa_data"

python process_absa_data.py

Training for TomBERT

• This is the training code of tuning parameters on the dev set, and testing on the test set. Note that you can change "CUDA_VISIBLE_DEVICES=6" based on your available GPUs.

sh run_multimodal_classifier.sh

Testing for TomBERT

• After training the model, the following code is used for directly loading the trained model and testing it on the test set

sh run_multimodal_classifier_test.sh

Implemented models

BERT and BERT+BL (run_classifier.py)

• You can run the following code to perform training and testing.

sh run_classifier.sh

TomBERT, mBERT, Res-BERT (run_multimodal_classifier.py)

• You can choose different models in the "run_multimodal_classifier.sh" file.

BERT and TomBERT trained by me

• You can download the BERT and TomBERT models trained by me. You can find the results we report in our paper from the "eval_result" files. https://drive.google.com/open?id=1e3rL3G1ojaDWZnrkmZX-uLudPbQo7tVe

Acknowledgements

• Most of the codes are based on the codes provided by huggingface: https://github.com/huggingface/transformers.