UniPoll: : A Unified Social Media Poll Generation Framework via Multi-Objective Optimization

The official implementation of the paper UniPoll: A Unified Social Media Poll Generation Framework via Multi-Objective Optimization.

This repository aims to automate the generation of polls from social media posts using advanced natural language generation (NLG) techniques. The goal is to ensure that even passive browsing users have their perspectives considered in text analytics methods.

Key Features:

• Automatic generation of polls from social media posts.
• Leveraging cutting-edge NLG techniques to handle noisy social media data.
• Enriching post context with comments to capture implicit context-question-answer relations.
• UniPoll framework: A novel unified poll generation approach using prompt tuning and multi-objective optimization.
• Outperforms existing NLG models like T5 by generating interconnected questions and answers.

Prepare the Environment

Please run the following commands to prepare the environment:

conda env create -f UniPoll.yaml
conda activate UniPoll
pip install git+https://github.com/tagucci/pythonrouge.git

WeiboPolls Dataset

The original data can be downloaded from this repo, you can also find them in data/WeiboPolls/origin.

Prepare the T5 checkpoint

Download the Chinese T5 checkpoint from here and put them under the checkpoints/imxly/t5-pegasus folder.

Experiments

Training

To reproduce the results in the paper, please run the following commands:

python finetuner.py configs/path_to_config_file.json

• You can find the config files in the configs folder, where configs are splitted according to different experiments. There is a detailed description of the correspondence between model names and configurations in the configs/README.md file.

• The processed data and results will be saved in the outputs/run_name folder.

• Please ref to config.py for more details about the configuration.

Inference

python inference.py

Reproduce the Main Results

If you want to reproduce the main results in the paper, please run the following commands:

export CUDA_VISIBLE_DEVICES=0
python finetuner.py configs/main_ablations/UniPoll.json

• After the training is finished, you can find the results in the outputs/UniPoll-t5 folder.

Results, Outputs and Checkpoint

• Here we provide the model outputs, results and checkpoint after executing the above commands. You can download them from here.

Results

This main results of poll generation are shown in the table below:

Model	ROUGE-1	ROUGE-L	BLEU-1	BLEU-3
UniPoll	47.92	45.02	39.96	22.78
T5	45.33	42.69	37.34	21.06
DUAL DEC	34.98	32.84	29.41	8.84
TOPIC	33.60	31.59	28.55	8.46
COPY	32.58	30.61	25.82	5.58
BASE	23.15	21.62	20.87	2.67

Demo

pip install gradio
python app.py --model_path "./outputs/UniPoll-t5/best_model" --device "cpu" 

Case Study

Citation

@misc{li2023unipoll,
      title={UniPoll: A Unified Social Media Poll Generation Framework via Multi-Objective Optimization}, 
      author={Yixia Li and Rong Xiang and Yanlin Song and Jing Li},
      year={2023},
      eprint={2306.06851},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}

Contact Information

If you have any questions or inquiries related to this research project, please feel free to contact:

• Yixia Li: liyixia@me.com