Manifesto-Anchored Method Maps Party Positions on Twitter

A team of NLP researchers has adapted party-position modeling from election manifestos to Twitter, using hashtags to fine-tune text representations and dropping the need for manual annotation. The paper, Toeing the Party Line, appeared in the Findings of EMNLP 2024, authored by Maximilian Maurer, Tanise Ceron, Sebastian Padó, and Gabriella Lapesa.

Prior work on political positioning has leaned on manifestos, the parties' formal electoral programs, because Twitter data is, in the authors' phrase, "ambiguous and often dependent on social context." The team extends a recently proposed method for predicting pairwise positional similarities between parties to the tweet setting, with hashtags carrying the supervisory signal in place of hand labels.

They report that the resulting positionings are "stable positionings reflective of manifesto positionings," holding up both when all tweets from candidates across years are available and when only smaller subsets from shorter time periods are. The paper's headline claim is that relative positioning can be recovered "without the need for manual annotation, even in the noisier context of social media."

The abstract does not name the countries, parties, or elections in the experiments, and no accuracy figures or baseline comparisons appear there. Two researchers on our tracked list surfaced the link, one small signal that the NLP-for-politics circle is passing it around.