[Paper Review] #IStandWithPutin versus #IStandWithUkraine: The interaction of bots and humans in discussion of the Russia/Ukraine war

Bridget Smart, Joshua Watt+3 more|arXiv (Cornell University)|Aug 15, 2022

Misinformation and Its Impacts4 citations

TL;DR

This study analyzes how bot and human accounts shape online discourse during the early phase of the Russia-Ukraine war using Twitter data (Feb 23–March 8, 2022). By combining sentiment analysis, linguistic profiling (LIWC), and information flow modeling, it reveals that pro-Russian non-bot accounts are the most influential overall, while bot activity significantly amplifies discussions on angst and governance, with bidirectional influence between bot activity and sentiment.

ABSTRACT

The 2022 Russian invasion of Ukraine emphasises the role social media plays in modern-day warfare, with conflict occurring in both the physical and information environments. There is a large body of work on identifying malicious cyber-activity, but less focusing on the effect this activity has on the overall conversation, especially with regards to the Russia/Ukraine Conflict. Here, we employ a variety of techniques including information theoretic measures, sentiment and linguistic analysis, and time series techniques to understand how bot activity influences wider online discourse. By aggregating account groups we find significant information flows from bot-like accounts to non-bot accounts with behaviour differing between sides. Pro-Russian non-bot accounts are most influential overall, with information flows to a variety of other account groups. No significant outward flows exist from pro-Ukrainian non-bot accounts, with significant flows from pro-Ukrainian bot accounts into pro-Ukrainian non-bot accounts. We find that bot activity drives an increase in conversations surrounding angst (with p = 2.450 x 1e-4) as well as those surrounding work/governance (with p = 3.803 x 1e-18). Bot activity also shows a significant relationship with non-bot sentiment (with p = 3.76 x 1e-4), where we find the relationship holds in both directions. This work extends and combines existing techniques to quantify how bots are influencing people in the online conversation around the Russia/Ukraine invasion. It opens up avenues for researchers to understand quantitatively how these malicious campaigns operate, and what makes them impactful.

Motivation & Objective

• To understand how bot-like accounts influence human discourse during the Russia-Ukraine war on social media.
• To quantify information flow patterns between bot and non-bot accounts across pro-Russian and pro-Ukrainian communities.
• To investigate the impact of bot activity on sentiment and linguistic content in online discussions.
• To assess whether malicious influence campaigns effectively shape public discourse during geopolitical conflicts.
• To develop a transferable framework for analyzing coordinated disinformation campaigns using linguistic, sentiment, and time-series techniques.

Proposed method

• Collected 5.2 million tweets containing specific hashtags related to pro-Russia and pro-Ukraine stances between February 23 and March 8, 2022.
• Classified accounts into bot-like (using Botometer) and non-bot groups, and further categorized by national lean (pro-Russian, pro-Ukrainian, balanced).
• Applied LIWC and VADER to extract linguistic features and sentiment scores from tweet content.
• Used time-series analysis and Granger causality to measure lagged effects of bot activity on sentiment and linguistic categories.
• Employed information theoretic measures to quantify net information flows between aggregated account groups.
• Conducted cross-correlation analysis between bot proportions and LIWC categories over 48-hour lags to validate directional effects.

Experimental results

Research questions

• RQ1 How do bot-like accounts influence the sentiment and linguistic content of non-bot accounts in the online discourse around the Russia-Ukraine war?
• RQ2 What are the patterns of information flow between bot and non-bot accounts, and how do they differ between pro-Russian and pro-Ukrainian communities?
• RQ3 Which linguistic and sentiment categories are most significantly influenced by bot activity, and with what temporal lag?
• RQ4 Is there a bidirectional relationship between bot activity and sentiment in non-bot discussions?
• RQ5 How do aggregated account groups (by national lean and bot status) compare in terms of self-entropy and influence capacity?

Key findings

• Pro-Russian non-bot accounts are the most influential overall, with significant net information flows to multiple other account groups.
• No significant outward information flows originate from pro-Ukrainian non-bot accounts, but strong inward flows exist from pro-Ukrainian bots to pro-Ukrainian non-bots.
• Bot activity significantly increases discussions on angst (p = 2.450 × 10⁻⁴) and work/governance (p = 3.803 × 10⁻¹⁸), with effects emerging after a 3–10 hour lag.
• A bidirectional relationship exists between bot activity and non-bot sentiment (p = 3.76 × 10⁻⁴), indicating mutual influence.
• Self-declared bots show the strongest impact on the 'Work' category (e.g., 'government', 'leadership'), with significant and sustained effects over 48 hours.
• Effects on 'Angst' and 'Filler' categories diminish within 24 hours, while 'Function' and 'Work' category effects persist up to 48 hours.