Detecting Characteristics of Cross-cutting Language Networks on Social Media
Rezvaneh Rezapour, Jaihyun Park, Jana DiesnerSocial networking sites offer people a space to discuss and share their points of view on controversial issues and topics of societal interest. As argued by Habermas, political homophily seems to be a hindrance when it comes to realizing the concept of the public sphere, where expressing, debating, and sharing opinions on controversial issues is conducive to political deliberation and decision. Although arguments around the benefits of cross-cutting communication (exchanging opinions across political lines) for enhancing political participation still seem unsettled, there is an agreement on the importance of being exposed to and considering different points of view in democracies. Analyzing cross-cutting communication has been a challenge in research related to opinion mining, and controversy detection. A large body of opinion mining research on societal issues is centered around classifying people’s stance (e.g., in favor of, against, or neutral) by leveraging machine learning and deep learning method, and enhancing the performance of the prediction models. However, we argue that an important step toward better understanding and predicting stance is to first extract and analyze the underlying aspects of the social issues (sub-topics) that are shared and discussed on online platforms. We believe that understanding these sub-topics can help in minimizing online segregation and bringing people with different points of view closer to each other.
Toward this goal, in this study, we explore the relationship between stance and cross-cutting communication on the sub-topic level and through the creation of word graphs. We leverage a benchmark stance dataset made available for SemEval 2016, which consists of 4,870 tweets labeled as “in favor”, “against”, or “neutral” toward six topics related to social or political events: abortion, atheism, climate change, feminism, Donald Trump, and Hillary Clinton. To extract sub-topics, we asked three annotators to extract 1 to 5 words that could best describe the aspects of the discussion in each tweet. After getting the labels, we disregarded the subtopics that were not shared by at least two annotators and constructed a word graph using the remaining ones. The sub-topic word graph represents the shared language and viewpoints, or the lack thereof, across different stances and topics. To further explore cross-cutting communication, we extract the related Twitter user network. Using this methodology enables us to address the following research questions:
(1) Does dividing broad concepts into subtopics improve stance prediction? (2) Is cross-cutting communication on a more granular level than broad concepts taking place? (3) Are the existence or lack of cross-cutting communication and related stance associated with differences in language?
Answering these questions will bridge a gap in understanding granularities of societal dialogue and potentially help to find ways to mitigate side effects of filter bubbles, echo chambers and communicative tribalism, and to promote cross-cutting communication. By understanding the network structure of the subtopics and different points of view, we expect to contribute to a public sphere in which deliberative democracy of the kind that Habermas had conceived could thrive.