Still out there: Modeling and Identifying Russian Troll Accounts on Twitter

Im, Jane; Chandrasekharan, Eshwar; Sargent, Jackson; Lighthammer, Paige; Denby, Taylor; Bhargava, Ankit; Hemphill, Libby; Jurgens, David; Gilbert, Éric

doi:10.1145/3394231.3397889

Cited by 61 publications

(66 citation statements)

References 22 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…This work has used thematic information, sentiment analysis, keyword/n-gram information, and syntactic n-grams (e.g. Boyd et al 2018;Im et al 2019;Ghanem et al 2019). While we have aimed at providing an overview of the linguistic properties of how troll data differ from genuine human accounts, we acknowledge that we simply do not yet have sufficient empirical evidence of the linguistic properties of troll data in English that originates from settings in which English is used as a non-native resource.…”

Section: Discussionmentioning

confidence: 99%

Twitter trolls: a linguistic profile of anti-democratic discourse

Lundberg

Laitinen

2020

Language Sciences

View full text Add to dashboard Cite

This article focuses on anti-democratic discourse and investigates the linguistic profile of Twitter trolls. The troll data consist of some 3.5 million messages in English obtained through Twitter in late 2018. These data originate from potentially state-backed information operations aimed at sowing discord in Western societies. The baseline data, against which the troll data are compared, contain circa 4.4 million messages in English drawn from the Nordic Tweet Stream corpus. A machine learning application that enables us to select genuine personal messages in this corpus is used to prune the data. The empirical part investigates frequencybased characteristics of the two datasets. We utilize a set of automatically-extracted word-list information and the observed frequencies of personal pronouns. Our empirical findings show considerable quantitative differences so that the troll data are shorter, make use of a smaller number of lexical types and tokens, and resemble more formal registers, while the personal messages are more spoken-like. The results could be used to improve automated detection systems whose purpose is to identify troll accounts.

show abstract

Section: Discussionmentioning

confidence: 99%

Twitter trolls: a linguistic profile of anti-democratic discourse

Lundberg

Laitinen

2020

Language Sciences

View full text Add to dashboard Cite

show abstract

“…Overall, our work is the most similar to Im et al (2019). In contrast to Im et al (2019), our work differs in two substantial ways.…”

Section: Introductionmentioning

confidence: 90%

“…Given ground-truth troll farm accounts, researchers have studied if they can develop classifiers to find other members of the troll farm organizations (Im et al, 2019). Even though all the accounts in their dataset are no longer active on Twitter (i.e., they have been banned), based on their classifier, they find that accounts with similar characteristics are still active.…”

Section: Introductionmentioning

confidence: 99%

How Many Users Are Enough? Exploring Semi-Supervision and Stylometric Features to Uncover a Russian Troll Farm

Nasrin¹,

Choo²,

Ko³

et al. 2019

Proceedings of the Second Workshop on Natural Language Processing for Internet Freedom: Censorship, Disinformation, and Propaga

View full text Add to dashboard Cite

Social media has reportedly been (ab)used by Russian troll farms to promote political agendas. Specifically, state-affiliated actors disguise themselves as native citizens of the United States to promote discord and promote their political motives. Therefore, developing methods to automatically detect Russian trolls can ensure fair elections and possibly reduce political extremism by stopping trolls that produce discord. While data exists for some troll organizations (e.g., Internet Research Agency), it is challenging to collect ground-truth accounts for new troll farms in a timely fashion. In this paper, we study the impact the number of labeled troll accounts has on detection performance. We analyze the use of self-supervision with less than 100 troll accounts as training data. We improve classification performance by nearly 4% F1. Furthermore, in combination with self-supervision, we also explore novel features for troll detection grounded in stylometry. Intuitively, we assume that the writing style is consistent across troll accounts because a single troll organization employee may control multiple user accounts. Overall, we improve on models based on words features by ∼9% F1.

show abstract

“…On the other hand, there are multiple approaches for detecting social bots, but they are unlikely to identify trolls [33], [34], this makes identifying political trolls automatically is still an open challenge [35]. To our knowledge only [34], [36], and [37] studied different features of trolls to build automatic identification approaches, all of them built their experiments solely on IRA dataset trolls revealed by Twitter.…”

Section: B Identifying State-sponsored Trollsmentioning

confidence: 99%

“…In [36], they developed three classifiers using Logistic regression, Decision Tree, and Adaptive Boosted Decision Tree. The troll accounts include 2286 out 3841 of IRA accounts disclosed by Twitter, only the accounts that used English as their main language were selected.…”

Section: B Identifying State-sponsored Trollsmentioning

confidence: 99%

Behavior-Based Machine Learning Approaches to Identify State-Sponsored Trolls on Twitter

Alhazbi

2020

IEEE Access

View full text Add to dashboard Cite

In recent years, there has been an increased prevalence of adopting state-sponsored trolls by governments and political organizations to influence public opinion through disinformation campaigns on social media platforms. This phenomenon negatively affects the political process, causes distrust in the political systems, sows discord within societies, and hastens political polarization. Thus, there is a need to develop automated approaches to identify sponsored-troll accounts on social media in order to mitigate their impacts on the political process and to protect people against opinion manipulation. In this paper, we argue that behaviors of sponsored-troll accounts on social media are different from ordinary users' because of their extrinsic motivation, and they cannot completely hide their suspicious behaviors, therefore these accounts can be identified using machine learning approaches based solely on their behaviors on the social media platforms. We have proposed a set of behavioral features of users' activities on Twitter. Based on these features, we developed four classification models to identify political troll accounts, these models are based on decision tree, random forest, Adaboost, and gradient boost algorithms. The models were trained and evaluated on a set of Saudi trolls disclosed by Twitter in 2019, the overall classification accuracy reaches up to 94.4%. The models also are capable to identify the Russian trolls with accuracy up to 72.6% without training on this set of trolls. This indicates that although the strategies of coordinated trolls might vary from an organization to another, they are all just employees and have common behaviors that can be identified. INDEX TERMS State-sponsored trolls, disinformation, propaganda, behavioral pattern.

show abstract

Still out there: Modeling and Identifying Russian Troll Accounts on Twitter

Cited by 61 publications

References 22 publications

Twitter trolls: a linguistic profile of anti-democratic discourse

Twitter trolls: a linguistic profile of anti-democratic discourse

How Many Users Are Enough? Exploring Semi-Supervision and Stylometric Features to Uncover a Russian Troll Farm

Behavior-Based Machine Learning Approaches to Identify State-Sponsored Trolls on Twitter

Contact Info

Product

Resources

About