We Don't Speak the Same Language: Interpreting Polarization through Machine Translation

Ashiqur R. KhudaBukhsh, Rupak Sarkar, Mark S. Kamlet, Tom M. MitchellAAAI 2021 · DOI

Introduction

In the current American political climate, it often feels as though opposing sides are speaking entirely different languages. Research presented at AAAI-21 by a team from Carnegie Mellon University formalizes this intuition. By applying unsupervised machine translation (MT) techniques to digital discourse, the authors demonstrate how political sub-communities use identical linguistic contexts to express diametrically opposed worldviews.

The Challenge of Linguistic Polarization

While partisans technically share the English language, the meaning and usage of specific terms have diverged sharply. Previous studies have identified "echo chambers," but quantifying the depth of this linguistic divide at scale remains difficult. Manual analysis of millions of posts is infeasible, and simple keyword counting fails to capture the nuance of how words function within a community's unique lexicon. The researchers address this by identifying "linguistic manifestations of polarization"—instances where different groups use distinct words to fill the same functional or emotional role.

Methodology: Politics as Machine Translation

The core innovation of this study is treating political polarization as a translation problem. The researchers modeled the comments of different news audiences (e.g., CNN vs. Fox News) as distinct "languages."

  1. Language Modeling: Using FastText, the authors generated word embeddings for each community, creating vector representations of "CNN-English," "Fox-News-English," etc.
  2. Manifold Alignment: Using neutral "anchor words" (standard English stopwords), the researchers aligned these vector spaces. In a typical MT task, this would align "apple" with "manzana"; here, it aligns words based on shared contextual roles.
  3. Identifying Misalignment: Once aligned, the system identifies words that do not translate to themselves. For example, if "solar" in the CNN corpus is mathematically closest to "fossil" in the Fox News corpus, it indicates these terms occupy the same "contextual slot" for their respective audiences.
Table 1: Examples of misaligned word pairs identified by the translation model across various categories.

Data and Scale

The study utilizes a massive longitudinal dataset of YouTube comments, comprising 86.6 million comments from 6.5 million unique users across over 200,000 videos. The data spans from 2014 through July 2020 and includes content from four major networks: CNN, Fox News, MSNBC, and OANN. Additionally, the authors analyzed 4 million comments from "Primetime Comedy" channels to determine which political "language" late-night fans tend to speak.

Key Findings: The Mirror Effect and Echo Chambers

The most striking result is the "Mirror Effect," where phrases like black lives matter and all lives matter are used in nearly identical linguistic structures. They serve as functional translations of one another within their respective echo chambers. The algorithm successfully identified several ideological mirrors without human intervention, such as (CNN) mask $\leftrightarrow$ (Fox) muzzle and (CNN) Anderson Cooper $\leftrightarrow$ (Fox) Sean Hannity.

To measure audience homogeneity, the authors developed a Viewership Disagreement Factor based on YouTube like/dislike ratios. They found that OANN has the lowest disagreement, suggesting an extremely ideologically homogeneous audience, while CNN shows the highest disagreement, indicating a more combative or diverse viewership.

Quantitative Results

By measuring the mathematical similarity between these "languages," the authors established a polarization spectrum: MSNBC $\rightarrow$ CNN $\rightarrow$ Fox News $\rightarrow$ OANN.

Table 2: Similarity matrix showing the percentage of linguistic overlap between different news network audiences.

As shown in the similarity metrics, MSNBC and CNN share a high degree of linguistic overlap. However, similarity drops significantly when comparing mainstream networks to OANN, with overlap falling to approximately 42% in some comparisons. Furthermore, viewers of political comedy were found to speak a language much closer to "CNN-English" (83.2% similarity) than "Fox-News-English" (75%).

Conclusion

By framing polarization as a translation problem, this research provides a powerful automated tool for decoding the diverging realities of the American public. While the study is currently limited to YouTube and word-level translations, it lays the groundwork for future unsupervised sentence-level analysis to capture more complex rhetorical strategies across various social media platforms.

Summary updated

BibTeX
@inproceedings{khudabukhsh2021we,
  author = {Ashiqur R. KhudaBukhsh and Rupak Sarkar and Mark S. Kamlet and Tom M. Mitchell},
  title = {{We Don't Speak the Same Language: Interpreting Polarization through Machine Translation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.1609/aaai.v35i17.17748},
}
Figure