Open jamesallenevans opened 4 years ago
This research shows an interesting example in combing psychological experiments and large-scale data to study the cause of online troll behaviors. Here are my questions:
(1) relationship between mood and social context: It seems that the researcher, to some degree, assumes an orthogonal relationship between mood and social context when designing the experiment (by treating them as two independent conditions). However, it is intuitive to think that negative social context could induce a bad mood and thus causing an increasing tendency of trolling behavior. Actually, in the "Anger begets more anger" section, the authors seem to indicate that participating in a trolling discussion could cause anger. Will this potential relationship between mood and social context influence the design/conclusion of this study?
(1) generalize to other social media platform Trolling behavior is also prevalent in social media platforms like twitter. In this kind of platform, it is hard to identify the social context as the posts are not organized in a single thread. How could this specific form of comment organization influence the trolling behavior? Will it form a certain segregation pattern (i.e., echo chamber) for trolls and normal people? Or will it be easier for normal people to get affected and conduct trolling behavior?
While this article seemed to heavily focus on methods of causal inference, I am left with a few questions regarding its statistics and assertions:
1.) How did they arrive at the numbers in the following statement: "Negative mood increases the odds of trolling by 89%, and the presence of prior troll posts increases the odds by 68%." The chart they present shows coefficients for negative mood, negative context and their interaction at 0.64, 0.52 and 0.41, respectively. I am wondering how they broke down the interaction coefficient between negative mood and negative content (because 0.89 = 0.64 + 0.25; 0.68 = 0.52 + 0.16; 0.41 = 0.25 + 0.16).
2.) Additionally, while looking at this chart, it is interesting that the (adjusted) R squared values were left out, so I wonder how much of their outcome variable was explained by their model. Also, I think feel like the standard errors seem pretty high in comparison to the coefficients, suggesting that their model is not the most precise which leads me to question the validity of their model.
3.) Regarding their analysis of CNN.com comments: the paper deemed any flagged comments as trolling comments, even after having experts hand label both non-flagged and flagged posts. While their recall is pretty high, their precision is much lower, showing that a fair amount of false positives in the flagged posts, i.e. a good amount of posts that were flagged were not in fact trolling comments. Would it better to incorporate a conditional probability, i.e. P(troll | post flagged), as opposed to assuming that if flagged, the post was trolling?
I found this article's discussion of mood and situational contexts that determine actors' trolling behavior very interesting and useful. However, I do have some issues with how the approach and findings can be applied to current social media participations, as a lot of content that uses recognizably ironic tone is no longer interpreted as trolling behavior, and the increasing spread of negativity might have also changed people's perceptions of negativity in online posts. What are the potential solutions to address such issues in this approach?
Regarding how trolling behavior was labeled as such, in the experimental part, this job was done manually by 2 experts, while in the large data analysis part (cnn.com) this was done by analyzing flagged comments.
Is it possible that the study may be influenced by the bias of the annotators in classifying the troll behaviors?
On page 4, the authors discuss the details of their experiments. I am not sure if I am convinced by the fact that they have their participants answer logical questions and word problems that vary in difficulties to stir positive and negative moods since they told the workers that their performance has no bearing on the compensation. Obviously, the paper shows that having done these exercises changed the mood of the participants, but would it be better if, for example, the researchers ask the participants to read positive/negative stories to better replicates real-life situation where people’s mood is affected by things they come across in their lives?
This paper focuses on the cause of the trolling behavior (user's mood and the surrounding discussion context) and follows a definition that behavior that falls outside acceptable bounds defined by this communities. I have a question here about how to identify the trolling behavior and troll-like comments as mentioned in the CNN online experiment mentioned in this article. How exclude the comments that may just like trolling comments seemingly?
@katykoenig: To answer your Q1, the logit model coefficients are log-odds. The true odds are computed using the exponential function.
So the true odds for Table 2 are:
This paper explains that trolling behavior is as much a group behavior as individual behavior, and is highly dependent on the discussion context as well as mood or experiences prior to the discussion. A potential question regarding the operationalization of trolling is how researchers define trolling as a behavior, and if that definition may evolve throughout time
(1) For a study I ran that involved manipulating participants' moods, the IRB review process was very stringent and we were required to incorporate a mood-boosting activity (for the participants in the negative mood condition), despite the fact that we debriefed all of our participants afterwards. Given that Cheng et al. didn't explicitly include a mood-boosting activity for participants in their neg mood or neg context conditions, I'm curious as to how they framed their proposal to the IRB (such that it got approved). (2) In the context of this paper: recall is what percentage of troll posts were recognized, and precision is how much of those recognized were actually troll posts. Recall = .94 and precision = .66, meaning that the coders were 'over predicting' troll posts (i.e., false positives). In the context of censorship, and bearing in mind that the authors' operationalization of trolling did not consider user intent: do you think it's ethically sound/safer to over censor in this way, or under censor?
Though the experimental design is rather convincing, I have one question about its operationalization. As a general rule, the participants should comment at least once in the experiment, which is not usual in reality. Couldn't this cause their unusual behavior as well as any bias on the estimation?
I like this study very much. This study used both experiment design and large-scale data analysis in an online discussion community, which solidly proved the situational impact of mood and previous discussion context on trolling behavior.
I am very interested in the experiment scenario ('universe') the author created. In the universe, workers could interact with the other workers assigned to the same space. This study scenario renders ‘quasi’ field observation of real interactions possible; however, at the same time, it also induces other confounding variables, for example, the interaction flow in each universe might be different across conditions. I am very interested in how we could balance this pros and cons and what kind of addition analysis we could perform to control these confounding variables.
While I like the two-stage setup of this study, I'm not entirely surely about the quality of data at the experimental stage. The participation metrics look pretty weak: an average of ~1.18 posts-per-participant, and barely 2 votes-per-participant. Additionally, the deviation in average quiz scores (11.2 vs 1.9) seems extremely large. Could this be attributed to low respondent motivation, given that it's Amazon MechTurk?
I'm not entirely certain how the authors controlled for pre-existing knowledge/bias at the experimental stage. 'Political affiliation' is too coarse a variable, especially if the posts used in the experiment were about the Clinton-Sanders standoff.
I'd be curious to know why a mixed-effects model was used here (because I haven't used one before). What are the advantages/reasons for assigning Users as random effects, and having the design parameters as fixed effects?
In a discussion about the causes of antisocial behavior, the authors make the hypotheses based on two aspects of mechanism. While the authors have claimed that the analysis was largely not causal nor external valid, I just consider the rules or interface in which the discussions take place would also play an important role in triggering anti-social behavior. A similar point is also mentioned by @ziwnchen. What if such a mechanism significantly affects the patterns of text data? How can a research design better corporates concerns like this, or
The results of this research seem to be intuitive that mode and online environment play a role in user's trolling behaviors. However, I do think there are some details I feel doubt about:
Using the CNN news website as the data resource is ok but insufficient. For me, I think trolling behaviors are more likely to happen in forums or chat groups (hard to collect info from). People are less involved in interactions when just commenting on online news from my experience.
One thing I am unsure about is the neg/pos mood stimulus. As mentioned in the article, money reward is guaranteed, so some task takers might not care much about the content of the quiz or even would not become more positive or negative after seeing the feedback of the quiz. Also, their original mood before taking the experiment might also different and impact the result.
Actually, test takers' mood might be changed when they see the previous posts, which means that he/she might feel positive before entering the discussion, but when faced with troll posts, he/she actually feel terrible before posting.
Another puzzle is from the use of the CNN dataset. I understand that in actual world, the situation is always more complicated, but I suspect that for some of the factors like time (mood), topics, previous posts (context) could pose cross-effects. The authors seem to neglect this part. Also, there are many outer elements unmeasured for these two mechanisms.
One of the interesting aspects of this paper is the poor precision but good recall for detection of troll posts in the CNN comments section. This means that the non-experts(at trolling detection) were over predicting (trigger happy) in flagging comments. It would be interesting to do this experiment in a controlled setting and have non-experts who have certain political orientations or view points on a particular topic (say, a binary topic like Abortion) and study the precision and recall of flagging for the two groups.
Like @ccsuehara mentions, the manual annotation was done by two experts. I wonder how the annotators were able to agree on the scope of trolling language and behavior? I also wondered if the reason for the non-expert overprediction was due to the nebulous operational definition given in the experiment and if in order to complete the task, the MTurkers veered on flagging more to be on the "safe side". Are there ways to control for such behavior either during the experiment or after? I would say that this might be a downside of MTurk, there is no interaction from the organizers that may help to clarify the experiment.
In looking at Fig. 2 graph B, we can see that the most trolling behavior occurred in Mondays/Sundays and the least on Saturdays. The authors link this temporal progression to the concept of mood -- I would be interested to see an extension of this paper in which this mood is further contextualized (or talk through how to provide further contextualization to this project), as the finding that mood is related to trolling indicates to me a significant larger social context.
Cheng et al 2017 discuss variables that influence online commentators' behavior and their likelihood of antisocial behavior such as trolling. It would be interesting to extend the study and to compare the study groups' behavior in online forums and the real world and let the participants discuss in a physical setting, e.g. a living room. It would be interesting to see if provoking claims evoke similar reactions or if they differ and to compare how a different setting might have an influence on the group dynamics.
I feel that the 'lingering' of trolling behavior is more important than other findings. Bad mood and prior trolling blogs cannot be independent. I would say the bad mood is usually caused by prior trolling comments and why people troll is because they feel bad (for the topic they are talking about, for 'personality', for venting real-life anger or whatever reasons).
I would rather take it as a community and look at it from a normative and cultural perspective. Each platform is different in its trolling behaviors and wordings, and how long the trolling stays because of factor A and B and C, for example.
From my perspective, people's mood is easy to be molded by the social text, and even manipulated. Take the recent coronavirus for example, in Chinese social media platforms, many people who have higher senses of empathy are even subject to vicarious traumatization, which is representative of social text's impact on our minds. So I am wondering the reasonability of setting mood and social text as the two dimensions in this research.
This paper intends to answer the question: "Can situational factors trigger trolling behavior?" One of the most intriguing findings in this paper is that initial posts in a discussion set a strong, lasting precedent for later trolling. This provides insight into the analysis and construction of the online discussion context. This further contributes to the argument that trolling is a situational behavior rather than individual behavior. One limitation fo the article is the analysis of effect of mood. This is due to limited understanding of signals of mood.
Post your questions here about:
Cheng, Justin, Michael Bernstein, Cristian Danescu-Niculescu-Mizil and Jure Leskovec. 2017. “Anyone Can Become a Troll: Causes of Trolling Behavior in Online Discussions.” WWW 2017: 1-14.