[1] Kumar, Srijan, et al. “An army of me: Sockpuppets in online discussion communities.” Proceedings of the 26th International Conference on World Wide Web. International World Wide Web Conferences Steering Committee, 2017.

Summary
This paper analyses various aspects of Sockpuppets and their behavior over online discussion communities through the study of a dataset of 62,744,175 posts and studying the users plus discussions within them. The authors identify and highlight various traits of sockpuppets and their pairs and then compare that with ordinary users to gain insight into how the sockpuppets operate. They also propose two models, one to identify sockpuppets from regular users and the other to identify pairs of sockpuppets in the online community.

Reflection
Widespread access to the internet and the ease with which one can create an account online, has encouraged many individuals to create sockpuppets and use them to deceive and drive online consensus. The scale of online communities which result in delayed repercussions for deceptive behavior has only encouraged individuals. If one has ever used social media, the chances are that one has been followed or mentioned by a suspicious looking account. Facebook and Twitter have millions of such fake sockpuppet accounts, and there are entire industries that are built around this. The paper brings to light some fascinating facts about the dynamics of sockpuppets in online communities and how can we identify and handle them. However, it leaves some questions unanswered.

Sockpuppeting coupled with the power of spambots and astroturfing have become powerful tools for organizations as well as states in some cases to manipulate public opinion and spread misinformation. One can come up with a system to flag such users and fake posts but when people are operating at such a high level of expertise, can such a system actually work? Even if we ban such accounts, it barely takes a few minutes to create a new account and come back online, how do we deal with this?

Twitter has an anti-spam measure, where it takes hashtags out of the trending section if the content of the tweets is irrelevant. While it sounds like a good measure, consider a scenario where an actual topic of concern is buried inside because of sockpuppets flooding Twitter with spam content over critical trending hashtags. So, the mechanism which is used to defeat spam is itself burying essential topics. How can we guarantee that such systems will adequately serve the purpose they are designed for? Also, in large-scale social media settings, do sock pockets actually exist in pairs?

Not only sockpuppets create a disturbance, but they also develop a sense of doubt amongst the ordinary user. Although people have grown accustomed to non-sense speaking accounts over social media, there has been a significant shift in trust on content published online. It has increased the credibility of fake news, while at the same time reduced the credibility of genuine news. This is very prevalent in the Indian political sphere. Disguised under the guise of IT-Cell (party sponsored organizations which are responsible for online campaigns), these groups use sockpuppets masquerading as other influential people to draw attention from essential topics. Follow comment threads [Example 1][Example 2].

From the technical point of view, models can be improved by using “Empath“[1], instead of LIWC. Empath is build using the word embedding models like word2vec and Glove and has a larger lexicon than LIWC. One problem with using unigram based features is that the model fails to capture the underlying meaning of the sentence. For example, for the model there is no difference between the two sentences “the suspect was killed by the woman” and “the woman was killed by the suspect.” Studies have also shown that deep learning based models perform significantly better than standard machine learning models especially in text/image classification[2]. Such complex models with advanced feature sets can be considered for effective labeling of posts.

In conclusion, although the paper highlights essential features to detect sockpuppets and proposes a model to identify them, sockpuppets have evolved to be more sophisticated and backed by technology. One must think of an efficient way to stop them at the source than to filter them after the damage is done.

References
[1] Fast, Ethan, Binbin Chen, and Michael S. Bernstein. “Empath: Understanding topic signals in large-scale text.” Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems. ACM, 2016.
[2] Bengio, Yoshua, and Yann LeCun. “Scaling learning algorithms towards AI.” Large-scale kernel machines 34.5 (2007): 1-41.

Read More

[1] Mitra, Tanushree, Graham P. Wright, and Eric Gilbert. “A parsimonious language model of social media credibility across disparate events.” Proceedings of the 2017 ACM Conference on Computer Supported Cooperative Work and Social Computing. ACM, 2017.

 

Summary

This paper proposes a model to quantify the perceived credibility of posts shared over social media. Using Twitter, the authors collect data on 1377 topics ranging over three months and constituting of a total of 66 million tweets. They study how the language used can define an event’s perceived credibility. Mechanical Turkers were asked to label the posts on a 5-point Likert scale which ranged from [-2 to 2]. The authors then went on to define four classes to classify credibility and defined 15 linguistic measures as indicators of perceived credibility level. The rest of the paper goes on to discuss the results and what can be learned from them.         

 

Reflections

There is a large scale penetration of social media in our daily lives. It has become a platform for people to share their views, opinions, feelings, and thoughts. Individuals use language which is frequently accompanied by ambiguity and figure of speech, which makes it difficult even for humans to comprehend it. With any event occurring around the world, tweets are one of the first things that start floating on the internet. This calls for a credibility check.

Use of Twitter is very apt because the language of Twitter is brief due to its character limitations, which makes it ideal to study language features. Although the paper performs a thorough analysis to create a feature set that can help quantify credibility, there are many features that can improve the model.

Social media is filled with informal language, which is hard to process from a natural language processing point of view. It is unclear how the model deals with it. For example, the word “happy” has a positive sentiment while the word “happppyyyyyyyyy” shows a more positive sentiment. The paper considers punctuation marks like question marks and quotations but fails to acknowledge a very important sentence modifier – the exclamation mark. It serves as an emotion booster. For example, the observe the difference between the sentences “the royals shutdown the giants” and “the royals shutdown the giants !!!!”.

Twitter has evolved over the years and with it the way people use it. Real-time event reporting tweets now span over multiple tweets, where each reply is a continuation of the previous tweet. Example. Tweets reporting news also include images or some sort of visual media to give a better idea of the ground reality. The credibility of the author and the people retweeting it also change the perceived credibility. For example, if someone with a high follower to following ratio makes a tweet or retweets some other tweet, its credibility will naturally increase. Can we include these changes to better understand the perceived credibility?

Subjective words like the ones associated with trauma, anxiety, fear, surprise, disappointment are observed to contribute to credibility. This raises the question, can emotion detection in these tweets contribute to perceived credibility? Having worked with emotional intelligence over twitter data, I believe we could come up with complex feature sets that consider the emotion of the tweet as well as the topic in hand and study how emotions play a role in estimating credibility.

One contradiction, I observed is that when hedge words like “to my knowledge” are used in the tweet, they contribute to higher perceived credibility. However, the use of evidential words like “reckon” result in lower perceived credibility, In regular language, both can be used interchangeably but evidently one increases credibility while the other decreases it. Why would this be the case?

There is one more general trend in the observations which is intriguing. In most of the cases, the credibility of a post is high if it tends to agree with the situation at hand. Does that mean, a post will have high credibility if it agrees with a fake event and have low credibility if it disagrees with it?

In conclusion, the paper does an exhaustive study of different linguistic cues that change the perceived credibility of the posts and discussed in detail how the credibility changes from one language feature to another. However, considering how social media has evolved over time, many new amendments can be made to this existing model to create an even more robust and general model.

 

Read More

[1] Cheng, Justin, Cristian Danescu-Niculescu-Mizil, and Jure Leskovec. “Antisocial Behavior in Online Discussion Communities.” Icwsm. 2015.

Summary

In this paper, Cheng et al. study three online discussion driven communities (CNN, IGN, and Breitbart) to understand the dynamics behind antisocial behavior in Online Communities. Their results are greatly derived from a quantitative study, where they inspect data to find trends in antisocial behavior and use the same to support their conclusions. This study also tracks how the activities of the users transform over time. The main deciding factor for an individual to be classified as an antisocial entity is the fact that they have been banned by moderators. The paper then goes on to explain a model to predict antisocial behavior using various features that they had discussed earlier.

 

Reflection

The paper answers many questions while leaving many unanswered. Although many conclusions seem intuitive to understand, it is amazing how simply going through the data answers the same questions. Like the one where the authors discuss “if excessive censorship causes users to write worse”. Intuitively, if one is punished for doing the right thing, the chances of repeating the same nice thing again reduces considerably.

What exactly is antisocial behavior? From one online community to another, this definition will change. There can not be a single defining line. For instance, 4chan users will tolerate more antisocial elements than users on Quora. Also, as we move from one geographical area to other, speaking habits will change. What is offensive and inappropriate in some culture might not be inappropriate in some other culture. So, what content is acceptable and to what extent?

Antisocial posts in this paper are labeled by the moderators. These are human moderators and their views are subjective. How can we validate that these posts are actually antisocial and not a positive criticism or some form of sarcasm? Secondly, on huge social networking websites which produce millions of posts every day how can moderation be translated at such a large scale? The paper provides four features and amongst them, the “moderator” feature has more weight in the classifier than the others. But with such large-scale networks, how can one rely on community and moderator features? The model also has a decent accuracy but when extrapolated to a large user base, it could result in banning of millions of innocent user accounts.

Coming to the technical side, the model shows relatively high accuracy during cross-platform testing using simple random forest classifiers and basic NLP techniques. While “Bag of words” model with random forest classifiers is a strong combination, they are insufficient to build the “post features”, in this case. Users have many different writing styles and much depends on the context in which words appear, so something more advanced than “bag of words” is needed. Word vectors would be a very good choice as they help capture context using the relative distance between two words. They can be easily tailored to the common writing style of the platform.

By taking, posts from the same user, we can build a sentiment index for each user. Sentiment index will help predict what the user, in general, feels about a particular topic and prevent incorrect banning. It is comparable to a browser keeping your search history to understand your usage patterns. One can also look at all posts from a general perspective and create an “antisocial index” for each post and only if the index is above a certain threshold, should the user be banned or be penalized. Penalties could include disabling users posting privileges for certain hours, so as to ensure that even if there is a false positive, an NBU is not banned.

In conclusion, the paper provides an informative and intriguing baseline to track antisocial behavior. Many techniques can be used to enhance the proposed model and create an autonomous content filtering mechanism.

Read More