Analysis of Nuanced Stances and Sentiment Towards Entities of US Politicians through the Lens of Moral Foundation Theory

Shamik Roy, Dan GoldwasserNAACL 2021

TL;DR

This paper introduces a novel approach to analyze the nuanced political stances and partisan sentiment of US politicians by leveraging Moral Foundation Theory (MFT). By applying deep relational learning to a large corpus of congressional tweets on gun control and immigration, the authors demonstrate that MFT can effectively explain political polarization beyond simple left-right divides and reveal how politicians frame issues differently when discussing specific entities.

What Problem Does This Paper Address?

Political discourse, especially on social media, is often characterized by polarization and the use of emotionally charged language. While traditional analyses often categorize politicians into binary "liberal" or "conservative" camps, this approach fails to capture the complexity of their stances on specific issues. Furthermore, understanding how politicians express sentiment towards particular entities (like specific laws, organizations, or groups) is crucial for a deeper comprehension of political communication. This paper tackles this by exploring how Moral Foundation Theory, which posits that human morality is based on five core foundations (Care/Harm, Fairness/Cheating, Loyalty/Betrayal, Authority/Subversion, Purity/Degradation), can be used to analyze these nuanced stances and entity-specific sentiments.

Methods

The core of this work lies in a deep structured prediction framework called DRaiL (Deep Relational Learning). This framework is used to identify moral foundations within tweets. DRaiL jointly models various features, including tweet text (encoded using BERT), author's political affiliation, and the topic of the tweet. It also incorporates temporal dependencies between tweets, treating them as soft constraints. This approach moves beyond simple lexicon matching or basic machine learning models by capturing complex relationships and dependencies within the data. For training and evaluation, the authors used a dataset of tweets from US politicians, augmented with a topic indicator lexicon built using Pointwise Mutual Information (PMI) scores.

Datasets

The study utilizes a dataset of approximately 161,000 tweets from US politicians on the topics of Gun Control and Immigration, collected between 2017 and February 2021. This dataset was built upon existing work by Johnson and Goldwasser (2018), which provided a smaller annotated subset. The authors also created topic indicator lexicons for Gun Control and Immigration using PMI scores. For evaluating the moral foundation identification model, a subset of 2050 tweets annotated for moral foundations was used, with a 5-fold cross-validation approach.

Key Insights and Ideas

  • MFT for Nuanced Stances: Moral Foundation Theory provides a powerful lens to understand political stances that go beyond the traditional left-right spectrum. The paper shows a strong correlation between the grades assigned to politicians by watchdog groups (representing nuanced stances) and their usage of specific moral foundations.
  • Polarization in MF Usage: Republicans and Democrats exhibit distinct patterns in their use of moral foundations, indicating a clear polarization in how they frame political issues. For instance, Republicans tend to use "Non-moral" and "Authority" foundations, while Democrats lean towards "Subversion" and "Harm" (on Gun Control) or "Loyalty" and "Cheating" (on Immigration).
  • Entity-Centric Sentiment: The analysis extends to how politicians express sentiment towards specific entities. The paper demonstrates that the same moral foundation can be used to convey different partisan sentiments depending on the entity being discussed, highlighting the importance of context. For example, "Harm" can be used positively by Republicans when discussing "police officers" but negatively by Democrats when discussing "migrant children."
  • Framing Effects: Even when using the same moral foundation, politicians from different parties can frame issues differently. For instance, both parties might use "Fairness" when discussing "Gun Owner" rights, but their specific arguments and emphasis will differ.

Key Results

The DRaiL model, incorporating all base rules and soft constraints, achieved the best performance in identifying moral foundations, with a weighted F1 score of 60.24% on the annotated dataset. The analysis of politician stances revealed significant correlations between their grades (A+ to F) and their moral foundation usage. For example, as politicians move from more conservative (Grade A) to more liberal (Grade F) stances, the usage of "Non-moral" foundations decreases, while "Harm" and "Loyalty" increase for Gun Control and Immigration, respectively. The study also identified entities with the highest disagreement in moral foundation usage between Democrats and Republicans, such as "Amendment" on Gun Control and "Donald Trump" on Immigration.

Limitations and Future Work

The authors acknowledge that while their model performs well, it does not outperform state-of-the-art MF classification results, particularly for classes with fewer examples (e.g., "Cheating," "Degradation"). They suggest that annotating more tweets for these low-frequency classes could improve overall performance. Future work could involve incorporating additional linguistic cues, behavioral aspects, and exploring the temporal evolution of sentiments towards entities in response to real-life events. The paper also suggests investigating the stances of politicians in the middle of the political spectrum (grades B-D) more qualitatively and extending the framework to other tasks like political framing analysis and misinformation detection.

Summary updated

BibTeX
@inproceedings{roy2021analysis,
  author = {Shamik Roy and Dan Goldwasser},
  title = {{Analysis of Nuanced Stances and Sentiment Towards Entities of US Politicians through the Lens of Moral Foundation Theory}},
  booktitle = {Conference of the North American Chapter of the Association for Computational Linguistics},
  year = {2021},
}
Figure