In [1], the work presented has a very strong argument for the need for language models in social computing platforms. This can be deconstructed using the following  two sections:

• SUMMARY: The paper first gives a theoretical base to the concepts that are used, along with a survey of related work. Here modality, subjectivity and the other linguistic measures used have been defined to capture the different perceived dimensions of a language model. The claims of all of them are warranted with the help of previous work. The statistical framework considered the problem as an ordered logistic regression one resulting in phrase collinearity (A common property in natural language expressions). The performance of the model is well documented with a sound defense for the validity. The overall accuracy of the model is a clear indicator of its use against the considered baseline classifiers. Implications are drawn on each of the defined measures based on the inferential statistical results of the model.
• REFLECTION: As a proponent for credibility level assessments of social media content, I favor the establishment of well-founded metrics to filter content. The paper is a strong step in the direction with a detailed account of the design process for a good linguistic model. The few immediate design opportunities that are regurgitated from the paper are:
  • The creation of a deployable automated system for content analysis adopting such a model. This can be a very interesting project where a Multi-agent Machine learning model using the CREDBANK system as its Supervised Learner, can help classify tweets in real time assigning credits to the source of the content. This will be monitored by another agent which reinforces the Supervised Learner, essentially creating a meta-learner.[2]-[5]
  • Adaptation of an ensemble of such models to form a global system which cross-verifies and credits information not just from a single platform but across multiple ones to give the metaphorical “global mean” as against the “local mean” in information. [6]
  • The model should account for the linguistic chaos even with newly created “Purrrrr” or “covfefe”. These lexiconic outliers could be captured with the use of Chaos Theory in Reinforcement Learning, which could be an entirely new avenue of research.

The paper also helped me understand the importance of capturing the different dimensions of a language model and corroborating it with evidence with tools of statistical inference.

[1]        T. Mitra, G. P. Wright, and E. Gilbert, “A Parsimonious Language Model of Social Media Credibility Across Disparate Events,” Proc. 2017 ACM Conf. Comput. Support. Coop. Work Soc. Comput. – CSCW ’17, pp. 126–145, 2017.

[2]        D. Li, Y. Yang, Y.-Z. Song, and T. M. Hospedales, “Learning to Generalize: Meta-Learning for Domain Generalization,” Oct. 2017.

[3]        F. Sung, L. Zhang, T. Xiang, T. Hospedales, and Y. Yang, “Learning to Learn: Meta-Critic Networks for Sample Efficient Learning,” Jun. 2017.

[4]        R. Houthooft et al., “Evolved Policy Gradients,” Feb. 2018.

[5]        Z. Xu, H. van Hasselt, and D. Silver, “Meta-Gradient Reinforcement Learning,” May 2018.

[6]        Marios Michailidis (2017), StackNet, StackNet Meta-Modelling Framework, URL https://github.com/kaz-Anova/StackNet

Read More

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.

This paper represents an in-depth and meticulous analysis of different linguistic and social network features that affect the credibility of a post. Mitra et. al. take intuition from linguistic cues such as subjectivity, positive emotions, hedging along with social network specifics such as retweets and replies to build a model that maps such features to the level of credibility. This study with thorough experimentation and validation not only provides strong evidence of the effects of such features but also gives qualitative insights and implications of the research.

The model fit comparison section specifically reflects several social network characteristics. For example, getting better explanatory power after including both the original texts and replies highlights the role of context and conversational nature of the social media interaction. Seeing the low predictive power of non lexicon based features, such as hashtags, caps and question marks, I am curious about whether all such features could be grouped into the “readability index” of the corpus corresponding to each event. It is possible that lower readability can be a good predictor of lower credibility. (Although, it is not clear just by intuition whether higher readability will be a good predictor of higher credibility)

Credibility in non-anonymous networks can have strong ties to how the source is viewed by the reader. Authors discuss that they did not include source credibility in the features but I think that “poster status” can also impact the perceived credibility. For example, I am more likely to believe in the fake news posted by my colleagues rather than a stranger with the same source. Similarly, I am more likely to believe in the information provided by a user with higher karma points than one with the lower karma points. Because the credibility annotations were done by turkers, it is not possible to assess the effect of poster status in the current setup. But, in a retrospective study, it is possible to have additional non-lexicon based features such as user statistics and tie strengths between the poster and the reader.

Such analysis that comprises of strong linguistic and non-linguistic features can be also applied to detecting fake news. Websites such as “Snopes”, “PolitiFact”  have pieces of news and the fact-check review on them tagged by “original content”, “fact rating” and “sources” which can be used either for stand-alone analysis or grouping the twitter event streams as fake or credible.

Finally, I believe that consequences of credibility range from disbelieving in scientific and logical information such as the importance of vaccinations and climate change to believing in conspiracy theories and propaganda.  Fast paced online interactions do not allow the users to analyze every piece of information they get. This makes the linguistic and social influence perspective on credibility more relevant and important in de-biasing the online interaction.

Read More

A Parsimonious Language Model of Social Media Credibility Across Disparate Events

Summary:

In this paper, the authors uncover that language used in tweets can indicate whether an event will be perceived as highly credible or less credible. After examining millions of tweets corresponding to thousands of twitter events they identify 15 linguistic features that can act as predictors of credibility and present a parsimonious model that maps linguistic constructs to different levels of credibility (“Low” < ”Medium” < ”High” < ”Perfect”). The events were annotated based on the proportion of annotations rating the reportage as ‘Certainly Accurate’. Authors use Penalised Logistic Regression for modeling and find that subjectivity is the most explanatory feature followed by positive & negative emotion. Overall, the results show that certain words and phrases are strong predictors of credibility.

Reflection:

As defined by Hiemstra et al in their paper “Parsimonious Language Models for Information Retrieval”, a parsimonious model optimizes its ability to predict language use but minimizes the total number of parameters needed to model the data.

All the papers we read for the last two reflections stressed the fact how linguistic constructs define the identity of a group/individual on online social communities. While the he use of racist, homophobic and sexist language was part of the identity of /b/ board in 4chan, users of a group in Usenet used “Geek Code” to proclaim their geek identity. We also learned how banned users on CNN used more negative emotion words and less conciliatory language.

I liked how authors validated their method of annotating the events with the HAC based clustering approach to group the events. They use Rand similarity coefficient to find the similarity between the two clustering techniques. The high R value indicates agreement between the two. I agree with author’s selection of annotation technique since it’s more generalizable.

Each mechanical turk needs to be aware of the event before annotating it. Otherwise they need to search for it online. How can can we ensure that the online news is not making the turker biased. Are turkers reading all the tweets in the pop up window before selecting a category or do they just base their decision by reading the first few tweets. I believe how an event is reported can greatly vary. So making a judgment by reading the first few tweets might not give a clear picture. Also, was the order of tweets in pop up window same for all the turkers? I believe I’ll find the answers to these questions after reading the Credbank paper.

The unit of analysis in this paper is an event rather than a tweet. And an event is perceived highly credible if large number of annotators rate the reportage as ‘certainly accurate’. But is the news perceived as credible actually credible? It will be interesting to see whether events perceived as credible are actually credible or not. A lot of work is going on in fake news detection and rumor propagation on social media platforms. Also, can people/organizations make use of this research to structure rumors in such a way that they are perceived credible? Will this reverse approach work too?

I believe a better way of calculating the value of “Questions” feature would be to calculate the proportion of tweets carrying question mark rather than counting the total number of question marks present in the tweets corresponding to an event.

One of the other features to determine credibility could be presence of URL in tweets. Specially URLs of trusted news agencies like CNN.

In the end I’ll reiterate the author and say that linguistic features combined with other domain specific features could act as foundation for an automated system to detect fake news.

Read More

Reading:

[1] A Parsimonious Language Model of Social Media Credibility Across Disparate Events

Summary:

This paper is geared towards understanding perceived credibility of information that is disseminated using social media platforms. In order to understand perceived credibility of published information, the authors of this paper decided to examine 66 million twitter messages a.k.a tweets, which were associated with 1,377 events that occurred over a period of 3 months — between October 2014 and February 2015. In order to examine these tweets from a linguistic vantage point, the authors came up with fifteen linguistic dimensions that assisted them to conceive a model that “maps language cues to perceived credibility.” In addition to the various linguistic dimensions the authors also highlighted the importance of particular phrases within these dimensions. To establish credibility of various tweets, the authors executed various experiments where the subjects were asked to rate a tweet on a 5 point Likert scale ranging from -2 (certainly inaccurate) to +2 (certainly accurate). The authors also employed nine control variables — in addition to the results from the experiment, linguistic dimensions, and identification of various phrases within these dimensions — that helped them account for the effect of  content popularity. The culmination of myriad of linguistic and statistical models ensued in a definitive parsimonious language model — howbeit the authors warn against independent usage of this model. Albeit they argue that the language model serves as an important step towards a fully autonomous system.

Reflection and Questions:

Comprehending the fact that utilization of specific words, writing styles, and sentence formations can alter the perceived credibility of a post made on social media surprises me, partly because its a new finding for me and partly because as an engineer I have never really paid attention to language formation but only to the facts in the text. Throughout the paper the authors have engaged in the idea of credibility of the post/tweet, howbeit according to my understanding it is the source of the information that necessitates “credibility” and text presented by the source necessitates “accuracy” and  “reliability”.  The authors write that “words indicating positive emotion were correlated with higher perceived credibility;” the question then arises: what about news bearing bad news? for instance death of a world leader; that news will not bear any “positive emotion.” Whilst reading the paper I came across a sentence stating that disbelief elicits ambiguity, which I disagree with. Disbelief can be used in a variety of combinations, none of which I think elicit ambiguity.

Reading the paper, I couldn’t help but think how does this model utilize slang language ? There could be a credible post that involves slang language because according to me millennial’s are more prone to trust a post that contains colloquial language instead of formal language, unless the source of information is associated with main stream media. The previous question alludes to the next question how is slang language in different countries/ usage of English in other countries taken into consideration ?  The reason I ask this is because specific words have different interpretations in different countries/ different regions of the same country resulting in different perceived credibility. As we are on the topic of interpretation of language in different regions, the question arises that: is this model universally suitable for all the languages in the world (with slight alterations), or would different languages require different models ? The main reason for this question is because people tweet in varied languages and language barrier could change perceived credibility of the post/tweet. Lets hypothesize that a post is originally made in a language other than English, howbeit English readers use the translate button on twitter/facebook to read the post, now the perceived credibility of post depends on the region the person resides in and the accuracy of translate feature of the particular social media platform. How can multiple considerations be synthesized to create a more suitable perceived credibility score for a specific situation ? 

 

Read More

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:

The authors of this paper examined how people perceived the credibility of events that were reported by users. The goal was to build a model that would be able to identify, with a good degree of accuracy, how credible a person will perceive an event based on the text of the tweets (and replies to the tweets) related to the event. To do this, they used a dataset of collected twitter streams with the tweets classified by events and the time when they were collected. These tweets were also rated for their credibility on a 5 point scale. The authors also put forward 15 language features that could be used to influence the credibility perception. With all this in hand, the authors were able to analyze and identify the words and phrases in the different language feature categories that corresponded to high and low perceived credibility scores for the various events. Though the authors advise against using the model on its own, it can be used to complement other systems such as fact checking systems.

 

Reflection:

What I found most interesting was the phenomenon of how the perception of credibility seems to flip for positive to negative and vice versa between the original tweets and replies. I first thought there might be a parallel here between tweets-replies and news article-comments but of course that doesn’t work because there are cases where the replies are perceived more credible than the original so that parallel doesn’t always work because the original tweets are not always credible. (Then again, there are cases where a news article is not necessarily credible so maybe there is a parallel here after all? I’m sorry, I’m grasping at straws here.)

“Everything you read on the internet is true. – Mahatma Gandhi.” This is a joke that you’ll sometimes see on Reddit but also serves as a warning against believing everything you read because you perceive it to be credible. The authors of this paper mentioned how the model can be used to identify content with low credibility and boot it from the platform before it can spread. But could it also be used by trolls or people with malicious intent to augment their own tweets (or other output) in order to increase the perceived credibility of their tweets? This could certainly cause some damage as we are talking about false information being more widely believed because it was improved thanks to algorithmic help where otherwise it may have had a low perceived credibility.

Another thing to consider is longer form content. This analysis is necessarily limited by their dataset which only consists of tweets. But I have often found that I am more likely to believe something if it is longer and articulate. This is especially apparent to me when browsing Reddit where I am more likely to believe a well written, long paragraph or a multi paragraph comment. I try to be skeptical but I still catch myself believing something because it happens to be longer (and also believe it more when it is gilded, but that’s for a different reflection). So the question that arises is: What effect does length have on perceived credibility? And how do the 15 language factors the authors identified affect perceived credibility in such longer form text?

 

Read More

The rise of online social platforms in the last two decades brought people from around the world closer together in an unprecedented manner. The forming of online communities created opportunities for spreading knowledge, initiating fruitful conversations, and openly expressing opinions among people from different geographical and social origins. With these online communities becoming larger, they face the inherent challenge of controlling undesirable social behavior.

Justin Cheng et al. address this challenge in their research: Antisocial Behavior in Online Discussion Communities by conducting a data-driven analysis of antisocial behavior on three different social platforms, then using insights gained from their analysis to develop a prediction model that can help the early detection of antisocial behavior. This work counts as a substantial contribution towards automating the process of identifying an undesirable online behavior.

Below are some facts that we could infer from this work:

1. A simple way of defining an undesirable behavior could be the actions that cause someone to be expelled from a community. Justin Cheng et al. consider users who are banned from posting or commenting on a social platform to be those who post undesirable or antisocial content. Of course, this could be subjective to a community, but it is still one of the most indicative acts of undesirability. 
2. Regarding selecting features for the prediction task, those features selected from the actions of community members and moderators have a more significant effect than those selected from the textual content. This could lead to two conclusions. It could strengthen the argument that undesirability is subjective to a community, but it could also be that selecting textual or content features to describe undesirable or antisocial behavior is a more challenging problem. But given that the prediction model performs relatively well across different platforms, we can conclude that antisocial acts could be reliably defined by the reaction of other community members.
3. The automation of detecting antisocial behavior in online communities could help moderators to better control the content, but it could not yet completely replace them. The involvement of a human to approve the prediction is necessary.

If I could have a chance to build over this work, I could focus on the following points:

1. Although the features selected from the reaction of the community are more descriptive of undesirable behavior, it has a drawback of being less effective for posts that are further in time than the time a user was banned, making it harder to make an earlier detection of an undesirable or antisocial behavior. Hence, improving the prediction through features selected from the content of the post could help to address this limitation.
2. As mentioned above, the automated detection of antisocial behavior could not yet completely replace the human decision. We can make use of this fact to even enhance the prediction by utilizing the decisions of human moderators to correct the prediction error and build a continuously improving model.

Finally, It is highly important and also challenging to control the spread of undesirable content in online social platforms. An undesirable content could range from simply being off-topic, to using swear words, to discrimination and bigotry, to spreading rumors and misinformation.

 

Read More

Cheng, J., Danescu-Niculescu-Mizil, C., & Leskovec, J. (2015). Antisocial Behavior in Online Discussion Communities. Proceedings of the Ninth International AAAI Conference on Web and Social Media Antisocial, 61–70. Retrieved from http://www.aaai.org/ocs/index.php/ICWSM/ICWSM15/paper/view/10469

The focus of the paper is categorizing anti-social behavior in online discussion communities. The inferential statistical approach taken by the paper by corroborating all claims with statistics is one that I appreciate. The approach in itself needs to be picked apart with a fine tooth comb to both understand the method followed and point a few discrepancies.

The paper claims to have adopted a “retrospective longitudinal analyses“. The long-term observational study in a subjects naturalistic environment is close to home as my current research hopes to study the “Evolution of trust”. A few key takeaways here are:

• The pool of study is limited to online discussion forums and not extended to general social media platforms. Since the author has not claimed the same or provided any evidence for the possibility of the same it is safe to say that this model is not completely generalizable. In platforms like Twitter where the site structure may be similar, the model adopted here might fail. A possible reason could be the option of retweeting on Twitter.
• The use of the propensity scores to determine causal effects by matching, according to my understanding, is a representational and reductional technique. It is representational because it considers a section of the data to represent all of it. Reductional because it discards a section of the data not used for the mapping. I wonder if this data loss has an impact on the outcome.
• The use of Mechanical Turk is always a good way to complete work that is not possible for Artificial Intelligence. In the above Human Intelligence Task the paper mentions the use of 131 workers with each post being averaged for three workers. The question that seemed important is whether this is required if a model is being built for another platform not covered by the one mentioned in the paper. As human hours can be expensive an alternative could be explored by compromising on the quality of the label classification and building a better model which will also make it more robust.
• The main question in the paper that I was hoping that was clearly answered but felt was not was “Can antisocial users be effectively identified early on?”. This can be a huge boon to have for any social media platform developer and/or designer. The promise of having very less or no trolls is like giving the customers a Charlies Chocolate Factory.

I wonder if this can be achieved by the introduction of an “Actor-Critic Reinforcement Learning algorithm“[1]. The use of a reinforcement learning algorithm lets the AI agent venture into the dark maze to find an exit. By rewarding the classification or flagging of a user in the right category we will be pushing it to train itself into becoming a good classifier of Anit-social behavior. The advantage of this model will be that the critic will ensure that the actor i.e. the agent performing the classification will not learn very quickly and will learn only the right things. It takes care of any anomalies that could occur. If the possibility exists then I feel this can be an area definitely worth pursuing through a course project.

REFERENCES:

[1] Konda, V. R., & Tsitsiklis, J. N. (2003). Actor-Critic Algorithms. Control Optim, 42(4), 1143–1166. https://doi.org/10.1137/S0363012901385691

Read More

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

In this day and age when “information overload” is widespread, the commodity everyone is eager to capture is attention. Users having the ability to do the same are sought after by companies trying to tout their next revolutionary product. But, there is one group of users which has particular ability to capture attention but the way they achieve it, make them, thankfully, undesirable to these establishments. These users through their vile and provocative mechanisms can send even the most civil of the netizens off their rails. But who and how these rogue actors function, can their behaviour be profiled at scale and then be used for nipping such bad actors from forums early from the bud? These are the questions [1] is trying to answer.

To start with, the paper distinguishes users from three websites namely CNN, IGN and Breitbart over a period of 18 months into two categories, Future Banned Users (FBUs) and Never Banned Users (NBUs). The FBUs are observed to have two subgroups, ones who concentrate their efforts to few threads or groups and ones who distribute their efforts to multiple forums. The author then measures the readability of the posts by these different categories of users to observe that the FBUs tend to have higher automated readability indices (ARIs) and displayed more negative emotions than NBUs. The author also measures the trend of user’s behaviour overtime to note any shift in their category label. The author later uses four different feature set namely post, activity, community and moderator to build a model to predict if a user will be banned or not.

To start with, the dataset is annotated by 131 workers in AMT. But due to the nature of selection of the workers nothing is known about the race, educational background or even political alignments which can definitely change the definition of “anti-social.” The diversity of opinion on what constitutes as “anti-social” is extremely important which the author hasn’t given much credence to.

Given the use of metric of using user deletions, effectiveness of such a model in forums where such user feedback mechanism is not present or in forums while such behaviour is norm and rampant, I believe, would be extremely low. What could be the metrics that could be adopted in forums like the ones mentioned? This could be an interesting avenue to explore.

Also, could these anti-social elements have a coordinated attack in-order take control over the platform? The individuals can bench members with more reports and and use its members who have less of these reports. The individuals can even create new accounts helping them steer a conversation to their cause. These are interesting strategies these individuals could adopt, the methods described in the paper would fail to detect. Can profiling these users’ content in order to ascertain their true identities create a slightly more robust model? This is something one can definitely try to work on in the future.

Another, interesting work that could be done is to identify the different methods through which these trolls try to elicit inflammatory behavior from their target. Also one could try to see how these mechanisms evolve, if they do, over time, as old ones tend to lose their ability to elicit such behaviour.

Can identifying users’ susceptibility in different forums or networks, be used to take preventive steps against anti-social behaviour? If one were to do that what are the different features that could be used to predict such susceptibility? Couple of features without much could be the number of replies the user gives to these trolls, the timespan the user has been active in the network and length of replies along with the sentiment. This if done could also be used to identifies trolls who have more sway over people.

Although, the intent and the motivation of the paper was excellent the content the paper left much to be desired.

Read More