Gaining and Losing Influence in Online Conversation

Gaining and Losing Influence in Online Conversation Arun Sharma and Tomek Strzalkowski ILS Institute, State University of New York – University at Albany 1400 Washington Ave, Albany, NY 12222 USA asharma5@albany.edu, tomek@albany.edu Abstract In this paper, we describe a study we conducted to determine, if a person who is highly influential in a discussion on a familiar topic would retain influence when moving to a topic that is less familiar or perhaps not as interesting. For this research, we collected samples of realistic on-line chat room discussions on several topics related to current issues in education, technology, arts, sports, finances, current affairs, etc. The collected data allowed us to create models for specific types of conversational behavior, such as agreement, disagreement, support, persuasion, negotiation, etc. These models were used to study influence in online discussions. It also allowed us to study how human influence works in online discussion and what affects a person’s influence from one topic to another. We found that influence is impacted by topic familiarity, sometimes dramatically, and we explain how it is affected and why. Keywords: Natural Language Processing, Computational Sociolinguistics, Social Computing, Influence Detection, Text Analysis, Data Analytics 1. Introduction This research was undertaken to help us understand how influential people behave in group conversations when they discuss the topics about which they have less knowledge. How does the topic of conversation affect their behavior in a group, specifically in an online conversation? Since a great deal of our communications occurs online, it is important to study and understand the sociolinguistic behavior of people who have high degree of influence in such environments. For this research, we recruited participants in groups of 3 to 5 to discuss topics that they are all familiar with. Once we identified the most influential participant in each group, we moved them to a new group of 3 to 5 participants that discussed a different topic that our influencers were less familiar with. We collected interactions from both rounds of discussions as our data set. We automatically analyzed the data for various sociolinguistic factors and computed the degree of influence for each participant, using the DSARMD toolkit (Broadwell et al., 2012). Some of the factors considered included the rate of topic introduction, participation frequency, as well as patterns of language use, dialogue acts, references to people and other named entities (Named Entity Tagging), etc. These factors were further aggregated into predictions of sociolinguistic behavior in conversation, such as Argument Diversity, Disagreement Measure, Network Centrality, Topic Control, Involvement etc. 2. Related Work Our objective is to study how people who are influential in one context retain or lose their influence in another context. We focused on the topic of conversation, and specifically the participants’ degree of interest and familiarity with it. Research on measuring influence is a relatively new area in computational linguistics and natural language processing; although it has some commonalities with automated sentiment analysis and dialogue understanding, which have much longer history. The key recent work that the present …

Content truncated to 500 words