Cross-Domain Label-Adaptive Stance Detection

Stance detection concerns the classification of a writer’s viewpoint towards a target. There are different task variants, e.g., stance of a tweet vs. a full article, or stance with respect to a claim vs. an (implicit) topic. Moreover, task definitions vary, which includes the label inventory, the data collection, and the annotation protocol. All these aspects hinder cross-domain studies, as they require changes to standard domain adaptation approaches. In this paper, we perform an in-depth analysis of 16 stance detection datasets, and we explore the possibility for cross-domain learning from them. Moreover, we propose an end-to-end unsupervised framework for out-of-domain prediction of unseen, user-defined labels. In particular, we combine domain adaptation techniques such as mixture of experts and domain-adversarial training with label embeddings, and we demonstrate sizable performance gains over strong baselines, both (i) in-domain, i.e., for seen targets, and (ii) out-of-domain, i.e., for unseen targets. Finally, we perform an exhaustive analysis of the cross-domain results, and we highlight the important factors influencing the model performance.

Setup

$ python3 -m venv ~/.virtualenvs/stance-detection
$ source ~/.virtualenvs/stance-detection/bin/activate

Updating project dependencies

# And to install the packages
$ pip install -r requirements.txt

Getting the datasets

• Stance Detection Benchmark Data (arc argmin fnc1 iac1 ibmcs perspectrum scd semeval2016t6 semeval2019t7 snopes)
• Will-They-Won't-They Data (wtwt)
• Emergent Data (emergent)
• Rumor has it Data (rumor)
• Multi-Target Stance Dataset Data (mtsd)
• Political Debates Data (poldeb)
• VAried Stance Topics Data (vast)

Format

The data files must be converted into json lines format:

{"uid": 1234, "label": 0, "hypothesis": "TEXT OF THE CONTEXT", "premise": "TEXT OF THE TARGET"}

The files for each dataset must be named with the following pattern: rumor_train.json, rumor_dev.json, rumor_test.json, where rumor should be replaced with the proper name of the dataset (see DATASETS below).

Running the code

Environment

DATASETS=(arc argmin fnc1 iac1 ibmcs perspectrum scd semeval2016t6 semeval2019t7 snopes emergent mtsd poldeb rumor vast wtwt)
TARGET=arc
datasets=("${DATASETS[@]/$TARGET}")

MODEL_NAME="roberta-base"
MAX_SEQ_LEN=100
LEARNING_RATE=1e-05

Label Embeddings Training

python src/stancedetection/models/trainer.py --data_dir "data/all/" \
        --model_name_or_path ${MODEL_NAME} \
        --output_dir ${OUTPUT_DIR} \
        --task_names ${datasets[@]} \
        --do_train \
        --do_eval \
        --model_type lel \
        --learning_rate ${LEARNING_RATE} \
        --weight_decay 0.01 \
        --per_gpu_train_batch_size 64 \
        --per_gpu_eval_batch_size 256 \
        --replace_classification \
        --num_train_epochs ${EPOCHS} \
        --warmup_proportion ${WARMUP} \
        --adam_epsilon 1e-08 \
        --log_on_epoch \
        --max_seq_length ${MAX_SEQ_LEN} \
        --evaluate_during_training \
        --gradient_accumulation_steps 1 \
        --seed ${SEED} \
        --fp16 \
        --cache_dir cache \
        --overwrite_output_dir

python src/stancedetection/models/trainer.py --data_dir "data/all/" \
        --model_name_or_path ${OUTPUT_DIR}/checkpoint-best \
        --output_dir ${EVAL_DIR} \
        --task_names ${TARGET} \
        --do_eval \
        --model_type lel \
        --per_gpu_eval_batch_size 256 \
        --max_seq_length ${MAX_SEQ_LEN} \
        --seed ${SEED} \
        --fp16 \
        --cache_dir cache \
        --overwrite_output_dir

Domain Adaptation

python src/stancedetection/models/trainer_da.py --data_dir "data/all/" \
        --model_name_or_path ${MODEL_NAME} \
        --output_dir ${OUTPUT_DIR} \
        --task_names ${datasets[@]} \
        --do_train \
        --do_eval \
        --learning_rate ${LEARNING_RATE} \
        --weight_decay 0.01 \
        --per_gpu_train_batch_size 64 \
        --per_gpu_eval_batch_size 256 \
        --replace_classification \
        --num_train_epochs ${EPOCHS} \
        --warmup_proportion ${WARMUP} \
        --adam_epsilon 1e-08 \
        --log_on_epoch \
        --max_seq_length ${MAX_SEQ_LEN} \
        --evaluate_during_training \
        --gradient_accumulation_steps 1 \
        --seed ${SEED} \
        --fp16 \
        --cache_dir cache \
        --overwrite_output_dir

python src/stancedetection/models/trainer_da.py --data_dir "data/all/" \
        --model_name_or_path ${OUTPUT_DIR}/checkpoint-best \
        --output_dir ${EVAL_DIR} \
        --task_names ${TARGET} \
        --do_eval \
        --per_gpu_eval_batch_size 256 \
        --max_seq_length ${MAX_SEQ_LEN} \
        --seed ${SEED} \
        --fp16 \
        --cache_dir cache \
        --overwrite_output_dir

References

Please cite as [1]. There is also an arXiv version.

[1] Momchil Hardalov, Arnav Arora, Preslav Nakov, and Isabelle Augenstein. 2021. Cross-Domain Label-Adaptive Stance Detection. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 9011–9028, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics.

@inproceedings{hardalov-etal-2021-cross,
    title = "Cross-Domain Label-Adaptive Stance Detection",
    author = "Hardalov, Momchil  and
      Arora, Arnav  and
      Nakov, Preslav  and
      Augenstein, Isabelle",
    booktitle = "Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing",
    month = nov,
    year = "2021",
    address = "Online and Punta Cana, Dominican Republic",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.emnlp-main.710",
    doi = "10.18653/v1/2021.emnlp-main.710",
    pages = "9011--9028"
}

License

The code in this repository is licenced under the CC-BY-NC-SA 4.0. The datasets are licensed under CC-BY-SA 4.0.