We investigate whether pre-trained bidirectional transformers with sentiment and emotion information improve stance detection in long discussions of contemporary issues. As a part of this work, we create a novel stance detection dataset covering 419 different controversial issues and their related pros and cons collected by procon.org in nonpartisan format. Experimental results show that a shallow recurrent neural network with sentiment or emotion information can reach competitive results compared to fine-tuned BERT with 20x fewer parameters. We also use a simple approach that explains which input phrases contribute to stance detection.

The paper can be found here.

@inproceedings{hosseinia-etal-2020-stance,
    title = "Stance Prediction for Contemporary Issues: Data and Experiments",
    author = "Hosseinia, Marjan  and Dragut, Eduard  and Mukherjee, Arjun",
    booktitle = "Proceedings of the Eighth International Workshop on Natural Language Processing for Social Media",
    month = jul,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.socialnlp-1.5",
    pages = "32--40",
    
}

The code is an extention of Hedwig implementation of BERT. Please follow Hedwig's instruction to install the requirements. (you do not need to install word embedding for this project)

Procon20 contains 419 different controversial issues with 6094 samples. Each sample is a pair of a (question, argument) that is either a pro(01) or a con(10). The dataset file can be found at data/ProconDual. Place dataset files in ../hedwig-data/datasets/ProconDual/

To train and evaluate the VADER-sent-GRU model on (train.tsv, dev.tsv, test.tsv):

python -m models.bert_lstm  --dataset ProconDual  --model bert-base-uncased --max-seq-length 256 --batch-size 8 --lr 2e-4 --epochs 30  --gpu 1 --early_on_f1 --seed 2035  --pooling

To train and evaluate the NRC-Emotion-GRU model on on (train.tsv, dev.tsv, test.tsv):

python -m models.bert_lstm_emotion  --dataset ProconDual  --model bert-base-uncased --max-seq-length 256 --batch-size 8 --lr 2e-4 --epochs 30  --gpu 1 --early_on_f1 --seed 2035  --max-em-len 11 --pooling --emotion-filters positive,negative