Stance detection

There are several types of stance detection tasks. “In-target” refers to the tasks where both the training and test datasets have the same target; “cross-target” refers to the tasks where models are trained on source targets and then evaluated on a dataset with different targets. Finally, zero-shot stance detection is where the targets can be anything and there is no specific training data.

There is also a variation in the type of targets: noun phrases (e.g., a person) or claims.

Aiyappa2024benchmarking

Reviews

• Alsaif2023review
• ALDayel2021stance
• Burnham2024stance - “a practical guide” for political scientists

Methods

• Allaway2020zero shot
• Hardalov2022few shot
• Zheng2022Stanceosaurus - multilingual
• Zhang2023investigating - chain of thought
• Kuo2024advancing - multimodal
• Zhao2024ZeroStance - using LLMs to generate synthetic dataset
• Akash2024can - LLMs’ performance on open-target

Datasets

A table from Alturayeif2023systematic : https://link-springer-com.proxyiub.uits.iu.edu/article/10.1007/s00521-023-08285-7/tables/3

• In-target
  • SemEval 2016 Task 6: https://aclanthology.org/S16-1003/
• https://paperswithcode.com/dataset/vast - features many targets
• P-Stance: Li2021P Stance

Specialized dataset

• Gyawali2024GunStance