Twitter and political culture: Short text embeddings as a window into political fragmentation

Mapping polarization and relationships in political discourse on
social media is challenging since politicians’ positions and relationships
can be hard to pin down. In this paper, we attempt to use
politicians’ tweets as a metric of their affinities using representation
learning, by modifying the Word2Vec method such that politicians
are directly encoded into a Euclidean space. Our analysis of Indian
politicians shows that the relatively populous, linguistically more
homogeneous northern states are cohesively clustered based on their
party affiliations, whereas southern states cluster based on geography.
We propose that computational methods can be useful in examining
the tensions of regionalist tendencies against dominant national
political narratives.