Detecting civil conflict and information biases in polarized environments in social media
Abstract
Specific Aim A. Develop a composite index of conflict intensity. Develop methods for detecting several conflict indicators in social media, to be combined into a composite index of conflict intensity. Use several measures of verbal and non-verbal user behavior and the associated network-scale effects, including: (1) a measure of divergence between polarized user clusters, in the network induced by what users share, repost, or like; this will be based on novel methods we propose for for polarized community detection, as well as on existing measures of modularity and boundary-based polarization; (2) a measure of alignment of the user communities induced by different polarizing issues; (3) a measure of the intensity of mutual aggression and hostile person-to-person sentiment (flame wars), as indicated both by sentiment and user commenting patterns in mixed commentary situations, i.e., when the members of polarized groups interact; (4) a measure of the vocabulary discrepancies between the opposing groups, including lexical meaning shifts and neologisms, represented by distributional patterns captured by vector space embeddings. (5) a measure of presence and propagation of opposing sentiment towards trigger topics and events in user-generated text. Specific Aim B. Analyze dynamic trends of the networks induced by user behavior. Detect and analyze the dynamics of polarized user networks, including user clique and cluster formation, their stability over time, and the changes in cluster modularity and density; track the formation and dynamics of user groups related to the conflict and individual user connections. B1. Polarized community detection. Test and compare existing measures such as modularity and polarity-at boundary, against novel methods to be developed in the scope of the project, including methods based on the multi-view graph representation of the user network and the adaptation of Bayesian topic modeling techniques representing users as probability distributions over topics, and topics as probability distributions over the content the users share, mention or like. B2. Assessment of ideological issue alignment. Analyze the dynamics of the user networks induced by the issues related to the particular conflict, representing the resulting space as a set of partially overlapping user clusters. Develop a measure of issue alignment using the overlap in user clusters. Specific Aim C. Flame war and political sentiment detection. C1. Measuring verbal aggression. The analysis of verbal user behavior, including flame war detection and identifying instances of inciting to action, including directed and person-to-person sentiment. C2. Measuring topic dominance. Analyze the trends in the number of users involved, their posting frequency, and topic distribution. C3. Detecting directed sentiment towards salient topics. Track polarity and intensity of the sentiment expressed within the opposing user clusters towards some salient topics, as represented by common hashtags and common collocates extracted from news provided by the news aggregators. We will also track the propagation of sentiment towards new events through the user network. C4. Tracking vocabulary shifts. Detect and measure vocabulary discrepancies between the opposing groups, including lexical shifts in meaning and introduction of neologisms. Specific Aim D. Detecting information biases. Develop methods that combine language-based and reaction-based bias detection. Develop the methods that use (1) collaborative filtering framework, (2) deep learning techniques using reaction-based data.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Sep 11, 2018
- Source ID
- W911NF1610174
Entities
People
- Anna Rumshisky
Organizations
- Army Contracting Command
- United States Army
- University of Massachusetts