English Social Media Stance Detection for Fact Checking

Encompassing English social media replies annotated for stance towards a claim, supporting a SemEval 2019 shared task. It is designed for the text classification task of predicting whether a reply supports, denies, questions, or comments on a claim. The dataset size is categorized as between 1,000 and 10,000 instances.

Use Cases

• Train a text classifier to predict the stance label (support, deny, question, comment) of a social media reply.
• Analyze linguistic patterns in reply text that correlate with different stance categories.
• Build a model for fact-checking pipelines by detecting denying or questioning replies to claims.
• Fine-tune a language model on the annotated reply and claim pairs for multi-class stance detection.

Strengths

• Data size is confirmed to be between 1,000 and 10,000 instances (1K<n<10K).
• Annotations were created via crowdsourcing, providing multiple human judgments.
• Specifically designed for a SemEval 2019 shared task, ensuring a clear benchmark.

Limitations

• Crowdsourced annotations may introduce label noise or inconsistency.
• Geographic bias is likely as the data region is tagged primarily as 'us'.
• The dataset is monolingual (English), limiting cross-lingual applications.

Provenance

Source

strombergnlp

Collection Method

Crowdsourced annotations for social media replies.

Time Range

Associated with the 2019 SemEval task.

Freshness

Last updated on 2022-10-25.

Geography

Primarily United States (Region:us).