Alt-Right Influence in YouTube comment sections regarding climate change discussion
Authors: Julius Brose, Franziska Rau, Robin Wenzel
In recent times there has been an observable correlation between climate change denial and the so-called alt-right movement or similar movements. The alt-right movement is a far-right, white nationalist movement which emerged in the last few years, starting from 2010. It is largely an online phenomenon and as thus an ideal subject for our research. For ease of understanding, we facilitate all similar movements under the term alt-right. As an online phenomenon, the alt-right movement uses their own platforms for communication and exchange, namely platforms such as Breitbart or BitChute, but are also commonly found on big platforms like YouTube to promote their theories and views, sparking controversies and trying to get new people into the movement.
Our Research focuses on identifying comments from alt-right members using machine learning methods. For this, we use the provided machine learning pipeline from the seminar in combination with a list of prominent dog whistles of the alt-right movement[1]. Dog whistles are terms or specific formatting used by these movements to identify themselves as members and find other members of the movement.
As the seminar is focused on the topic of climate change and denial of climate change, we focus our research on alt-right influence in the comment sections of these kinds of videos. As we mentioned in the introduction, there is a lot of activity from the alt-right movement regarding these videos, as they mostly present a strong anti-climate-change stance. At the time of writing there is no list of specific videos we will use, the identification process of suitable videos will be part of the research project. The raw comment data will be accessed using the provided API of YouTube. We will not use comment data from alt-right (video-) platforms as basis for our research as these comments are lacking in diversity in the voiced opinions. They are for the most part only containing alt-right opinions, as non alt-right people rarely comment on or even use these platforms. Also, the alt-right members do not really need to use dog whistles on these platforms to identify themselves as members of the movement as they are in “safe space” of their movement and it is generally assumed that you belong to the movement when you are participating there.
We use the provided tool from the seminar to convert the comment data in an appropriate format for the machine learning pipeline to consume the data and produce clusters. The clusters the pipeline returns as a result are compared to the list of dog whistles and we try to find terms from the list which can be used to label the clusters. This means that we examine if the comments in the cluster are all similar or relevant to a dog whistle from the list or even all contain the same specific dog whistle. Regarding these points, our specific research question can be formulated as following: To what extent can a language model be used to identify and analyse dog whistles?. The answer to this question will be based on how many fitting labels for the clusters can be derived from the list of dog whistles. Specific numeric values for a hit rate are currently not set by us and will emerge when the research progresses further.
We try to evaluate what changes in the pipeline could be useful to optimize the identification of these dog whistles. Changes on parameters for the algorithm like the number of clusters can heavily influence the results and may lead to a more accurate identification of clusters fitting to a specific dog whistle. We also want to look into if this approach could be used to identify single users of YouTube as alt-right and possibly sanction them. As a full evaluation regarding this topic would be of a too large scale for this seminar and also would include using non-anonymized data, potentially putting the researchers or the university at risk of legal problems, we will focus on a theoretical, high-level analysis of this possibility. Our results could then possibly be used as a starting point for future work by other groups working in a different context, enabling them to work more freely on this.
Our analytical approach is a form of grounded theory, as we build the theory that the pipeline will be able to identify these dog whistles based on our previous results from using the pipeline on YouTube comment data and the resulting clusters. In our research we then try to prove this theory with our specific results.
Alt-Right Influence in YouTube comment sections regarding climate change discussion Authors: Julius Brose, Franziska Rau, Robin Wenzel
In recent times there has been an observable correlation between climate change denial and the so-called alt-right movement or similar movements. The alt-right movement is a far-right, white nationalist movement which emerged in the last few years, starting from 2010. It is largely an online phenomenon and as thus an ideal subject for our research. For ease of understanding, we facilitate all similar movements under the term alt-right. As an online phenomenon, the alt-right movement uses their own platforms for communication and exchange, namely platforms such as Breitbart or BitChute, but are also commonly found on big platforms like YouTube to promote their theories and views, sparking controversies and trying to get new people into the movement.
Our Research focuses on identifying comments from alt-right members using machine learning methods. For this, we use the provided machine learning pipeline from the seminar in combination with a list of prominent dog whistles of the alt-right movement[1]. Dog whistles are terms or specific formatting used by these movements to identify themselves as members and find other members of the movement.
As the seminar is focused on the topic of climate change and denial of climate change, we focus our research on alt-right influence in the comment sections of these kinds of videos. As we mentioned in the introduction, there is a lot of activity from the alt-right movement regarding these videos, as they mostly present a strong anti-climate-change stance. At the time of writing there is no list of specific videos we will use, the identification process of suitable videos will be part of the research project. The raw comment data will be accessed using the provided API of YouTube. We will not use comment data from alt-right (video-) platforms as basis for our research as these comments are lacking in diversity in the voiced opinions. They are for the most part only containing alt-right opinions, as non alt-right people rarely comment on or even use these platforms. Also, the alt-right members do not really need to use dog whistles on these platforms to identify themselves as members of the movement as they are in “safe space” of their movement and it is generally assumed that you belong to the movement when you are participating there.
We use the provided tool from the seminar to convert the comment data in an appropriate format for the machine learning pipeline to consume the data and produce clusters. The clusters the pipeline returns as a result are compared to the list of dog whistles and we try to find terms from the list which can be used to label the clusters. This means that we examine if the comments in the cluster are all similar or relevant to a dog whistle from the list or even all contain the same specific dog whistle. Regarding these points, our specific research question can be formulated as following: To what extent can a language model be used to identify and analyse dog whistles?. The answer to this question will be based on how many fitting labels for the clusters can be derived from the list of dog whistles. Specific numeric values for a hit rate are currently not set by us and will emerge when the research progresses further.
We try to evaluate what changes in the pipeline could be useful to optimize the identification of these dog whistles. Changes on parameters for the algorithm like the number of clusters can heavily influence the results and may lead to a more accurate identification of clusters fitting to a specific dog whistle. We also want to look into if this approach could be used to identify single users of YouTube as alt-right and possibly sanction them. As a full evaluation regarding this topic would be of a too large scale for this seminar and also would include using non-anonymized data, potentially putting the researchers or the university at risk of legal problems, we will focus on a theoretical, high-level analysis of this possibility. Our results could then possibly be used as a starting point for future work by other groups working in a different context, enabling them to work more freely on this.
Our analytical approach is a form of grounded theory, as we build the theory that the pipeline will be able to identify these dog whistles based on our previous results from using the pipeline on YouTube comment data and the resulting clusters. In our research we then try to prove this theory with our specific results.
[1] https://rationalwiki.org/wiki/Alt-right_glossary, last accessed: 07.01.2021