Conversational Inequality Through the Lens of Political Interruption

Clay H. Yoo, Jiachen Wang, Yuxi Luo, Kunal Khadilkar, Ashiqur R. KhudaBukhshIJCAI 2022 · DOI

TL;DR

This paper introduces a novel, large-scale dataset of political dialogues annotated for interruptions, enabling a nuanced analysis of conversational inequality. They found that while men tend to speak more words per line, women interrupt more frequently, and political polarization significantly influences interruption patterns, particularly in news media.

What Problem Does This Paper Address?

Political polarization in the US has been extensively studied, but its manifestation in conversational dynamics, specifically through interruptions, remains less understood. Interruptions can signal dominance, disagreement, or a desire to contribute, and their patterns can reveal underlying power imbalances and social dynamics. Existing research on interruptions often relies on small, manually annotated datasets, limiting the scope and generalizability of findings. This paper aims to bridge this gap by creating a large-scale dataset of political conversations from major US cable news networks and analyzing interruption patterns across various demographic and political dimensions.

Methods

The authors developed a novel dataset of over 625,000 interruption dialogues extracted from CNN, Fox News, and MSNBC transcripts. They then manually annotated a subset of 2,000 dialogues with fine-grained interruption labels, categorizing them into five types: floor-taking, overlap, disagreement, cooperative, and topic-changing. For broader analysis, these were further grouped into "benign" (overlap, cooperative) and "intrusive" (floor-taking, disagreement, topic-changing) categories. They trained a machine learning classifier (BERT-based with GRU layers) to predict interruption types, achieving a weighted F1 score of 82.4% for the binary benign/intrusive classification. This classifier was then used to analyze interruption patterns across different speakers and media outlets.

Datasets

  • Primary Dataset: A collection of 625,409 interruption dialogues extracted from 275,420 transcripts from CNN, Fox News, and MSNBC spanning January 2000 to July 2021.
  • Annotated Dataset (INTERRUPT): A subset of 2,000 dialogues from the primary dataset, manually annotated with fine-grained interruption labels.
  • Speaker Annotation: 1,666 frequently appearing speakers from the dataset were annotated for gender, race/ethnicity, occupation, and political orientation.

Key Insights and Ideas

  • Gendered Interruption Dynamics: While men tend to speak more words per line, women interrupt more frequently overall. This suggests a complex interplay where women might assert their presence through more frequent interruptions, even if men dominate in sheer volume of speech.
  • Political Polarization Amplifies Intrusive Interruptions: Individuals tend to use more intrusive interruptions when speaking with those of differing political beliefs. This effect is particularly pronounced in news media, where partisan bias seems to influence how journalists interrupt speakers with opposing political views.
  • Journalists as Mediators (and Interrupters): Journalists, particularly those in talk show host or anchor roles, interrupt more frequently, likely to manage the flow of conversations. They also show a higher proportion of benign interruptions, suggesting a role in facilitating dialogue.
  • News Media Bias and Interruption Styles: The study found that journalists from CNN and MSNBC tend to use more intrusive interruptions when speaking with Republicans, while Fox News journalists use more intrusive interruptions when speaking with Democrats. This highlights how media bias can shape conversational interaction patterns.

Key Results

  • Classifier Performance: The binary classifier for benign vs. intrusive interruptions achieved a weighted F1 score of 82.4%.
  • Gender Differences: Female speakers had a higher overall interruption rate (39.4%) compared to male speakers (35.9%). Male speakers tended to use more intrusive interruptions when speaking to other males (61.3%) than to females (54.3%).
  • Political Orientation: Speakers used more intrusive interruptions (74.8% for Democrats vs. Republicans, 70.5% for Republicans vs. Democrats) when conversing with individuals of opposing political views.
  • Media Bias: CNN and MSNBC journalists showed higher intrusive interruption rates when speaking with Republicans, while Fox News journalists showed higher intrusive rates when speaking with Democrats.
  • Occupation: Journalists interrupted more (39.0%) than speakers with political occupations (34.2%), with journalists also showing a higher ratio of benign to intrusive interruptions.

Limitations and Future Work

  • Selection Bias: The study is limited to speakers from major US cable news networks, who are likely to be highly educated and hold positions of social status. This may lead to a resistance to certain biases and might not generalize to broader populations. The focus on US politics also limits cross-cultural generalizability.
  • Annotation Granularity: The authors acknowledge that their binary classification of benign/intrusive interruptions might not fully capture the nuances of dominance, as some "benign" overlaps could still indicate power dynamics. Future work could explore more fine-grained interruption classification.
  • Successful vs. Unsuccessful Interruptions: The current analysis does not distinguish between successful and unsuccessful interruptions, which could be a factor in understanding dominance.
  • Non-Content Cues: The analysis is based solely on textual transcriptions, missing crucial non-content cues like intonation, pitch, and body language, which can significantly influence the interpretation of interruptions.
  • Future Directions: The authors suggest their dataset can be used to improve Question Answering systems and predict interruptions in group conversations. They also propose using the data to understand sociopolitical dynamics and raise awareness about gender and racial disparities in news media.

Summary updated

BibTeX
@inproceedings{yoo2022conversational,
  author = {Clay H. Yoo and Jiachen Wang and Yuxi Luo and Kunal Khadilkar and Ashiqur R. KhudaBukhsh},
  title = {{Conversational Inequality Through the Lens of Political Interruption}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2022},
  doi = {10.24963/ijcai.2022/724},
}
Figure