Thank you very much for your interest in this work.

This project was completed during my internship at Meituan over two years ago and is no longer being maintained.

I apologize for any inconvenience.

MuSE

We provide our collected dataset (MT-Product) in anonymous Google Drive.

And the other public datasets can be downloaded from:

• twitter15 and twitter17

• mvsa-single and mvsa-multiple.

Setup

This implemetation is based on Python=3.7. To run the code, you need the following dependencies:

• torch==1.11.0
• transformers==4.18.0
• torchvision==0.12.0
• spacy==2.3.7
• scikit-learn==1.0.2
• tensorboard==2.8.0

Or you can simply run:

pip install -r requirements.txt

Repository structure

|-- code
    |-- main.py  # the main function to run the code
    |-- utils.py # the preprocess for all datasets
    |-- models.py # the key models in our experiments, including CrossTransformer
    |-- run_example.sh  # the run example
|-- data
    |-- MT-Product # the example data for MT-product texts. For the full data, please download from the google drive in the above.
|-- requirements.txt

Run example

We give the run example of twitter15 dataset:

TASK_NAME="twitter15"
alpha=0.0001
beta=0.0001
theta=0.1
sigma=1.0
replace_start=1
replace_end=3
cls_init=0
num_layers=6
crf_dropout=0.5
learning_rate=1e-4
crf_learning_rate=1e-4
bert_type='uncased'
cross_dropout=0.2

CUDA_VISIBLE_DEVICES=0 python -u main.py \
    --task ${TASK_NAME} \
    --per_gpu_train_batch_size 40 \
    --per_gpu_eval_batch_size 40 \
    --alpha ${alpha} \
    --beta ${beta} \
    --theta ${theta} \
    --output_dir ../outputs/${TASK_NAME}_output/alpha${alpha}_beta${beta}_theta${theta}_sigma${sigma}_rs${replace_start}_re${replace_end}_cls${cls_init}_l${num_layers}_lr${learning_rate}_clr${crf_learning_rate}_${bert_type}_cd${cross_dropout}_last/\
    --do_train \
    --do_eval \
    --num_train_epochs 10 \
    --logging_steps 100 \
    --save_steps 100 \
    --evaluate_during_training \
    --num_workers 8 \
    --learning_rate ${learning_rate} \
    --crf_learning_rate ${crf_learning_rate} \
    --num_layers ${num_layers} \
    --replace_start ${replace_start} \
    --replace_end ${replace_end} \
    --cls_init ${cls_init} \
    --crf_dropout ${crf_dropout} \
    --skip_connection \
    --use_quantile \
    --bert_type ${bert_type} \
    --cross_dropout ${cross_dropout}

You can change the 'TASK_NAME' to run other datasets. Also, you can try any parameters in the python scripts.

Notice: you need to change all 'your url' in 'main.py' to the real url of datasets and models.

CrossTransformer

Here is the core of CrossTransformer. For more details, please refer to the 'models.py'.

def _cr_block(self, x1, x2, attn_weight1, attn_weight2):
        cls_weight1 = attn_weight1[:, 0, :]
        cls_weight2 = attn_weight2[:, 0, :]
        x1_mean = torch.mean(x1, dim=-2)
        x2_mean = torch.mean(x2, dim=-2)
        for i in range(cls_weight1.shape[0]):
            if self.use_quantile:
                theta1 = np.quantile(cls_weight1[i][1:].detach().cpu().numpy(), self.theta)
                theta2 = np.quantile(cls_weight2[i][1:].detach().cpu().numpy(), self.theta)
            else:
                theta1 = self.theta
                theta2 = self.theta
            
            for j in range(1, cls_weight1.shape[1]):  # except the first token, namely [cls]
                if cls_weight1[i][j] < theta1:
                    x1[i][j] = x2_mean[i] + x1[i][j] if self.skip_connection else x2_mean[i]
                if cls_weight2[i][j] < theta2:
                    x2[i][j] = x1_mean[i] + x2[i][j] if self.skip_connection else x1_mean[i]
        return x1, x2

Attribution

Parts of this code are based on the following repositories:

• UMT

• NIC

• PixelCNN++