Digital democracy Germany

Election monitor Germany 2021 - Research brief #2

Read this in German here. 

In this month’s research brief, we turn to look at the words used in Germany’s 2021 federal elections. Particularly, at the words used in the comments sections of YouTube when discussing the main chancellor candidates.

Key takeaways

  • Annalena Baerbock receives the most toxic YouTube comments.

  • The AfD is perceived quite favourably in YouTube comments and has many niche channels with active engagement.
  • Although Baerbock is attacked on the basis of her gender quite harshly (i.e. her competence questioned, sexualized language), the same language patterns have not been observed with Janine Wissler of Die Linke or Alice Weidel of the AfD; in fact, Weidel is described using quite positive terms (i.e. intelligent, sincere).  

  • Even if a video discusses policy and political agendas, many YouTube comments contain personal attacks, many focusing on personal scandals.

  • Personality is getting more attention than politics: Videos on candidates receive twice as many comments and more negative sentiment than videos about the election in general. 

Dashboard infographics

New things for you to explore

To mark this issue's analysis of words on YouTube, together with our partners at Der Tagesspiegel, we have added two new features to our social media dashboard: the most common words used by and about the candidates on Facebook, Twitter and Instagram - and a section where you can find the most liked, viewed and commented upon YouTube videos about the candidates.

Most common words about candidates (Twitter, Facebook, Instagram)

Popular YouTube videos about candidates

Infographic exploration

The trolls come out at night

Here, you can see that toxic comments and insults peak from midnight to 5 am, with a little peak during the afternoon commute when overall commenting hits its peak. 

However, as the number of comments increases during the day, we also see that they overall become more positive. This indicates that the larger community is not taking the bait of the trolls.


Deep dive

Hot and nasty 

A recent report in Der Spiegel found that the Greens’ candidate Annalena Baerbock is subject to a large number of hate comments in right-wing Facebook groups. While interesting and important, we want to examine the conversation in a more open online setting. This is where we are all exposed to the discussion and where people can be influenced and polarised by content and comments. 

Let’s look at how the different candidates are treated in YouTube video comments by looking at their sentiment, their toxicity, and the relationships of the words used to one another. 

Our key observations

  • The overall mood of comments on the election is rather neutral, not bad, not great. However, there are notable spikes of negative sentiment driven by events (e.g. the recent flood disaster in parts of Germany). 

  • Conversely, videos that are more candidate-centric receive mostly negative comments. This negative mood is fairly evenly spread aside from Alice Weidel and Christian Lindner. We believe this mood indicates not so much whether a candidate is subject to attack in a comment, but also the mood of their supporters. AfD and FDP supporters seem optimistic, in their own unique ways. 
  • Annalena Baerbock receives the most toxic comments. Attacks based on a person’s identity are less common, as are threats, but in this area, Tino Chrupalla is on the receiving end most often. References to his ignorance stand out (perhaps due to his lower educational level than the rest of the field). 

  • The toxic comments about the Green’s candidate stand out further when looking at word associations. For Baerbock, these included not only the sadly all too common attacks on competence and leadership, but also sexualised and derogatory terms. Neither of the other female candidates sees such treatment to a measurable extent. 

  • The AfD, on the other hand, is perceived quite favourably in YouTube comments. They seem to have many niche channels with active engagement – channels that those who oppose the AfD are unlikely to seek out and comment upon. There is a fair amount of ‘spirited’ and combative words, which likely drives the positive mood we noted above.

Why YouTube comments? 

YouTube is perhaps the most level of all major social media platforms, as most candidates do not have their own channels while party channels are not big players. Thus, the content is about them rather than by them, allowing for a more insightful look into who is being talked about and how. Being talked about on YouTube matters, as videos from here are particularly prone to being shared across platforms. In addition, YouTube has not been prominently covered in social media monitoring during the last German elections and is thus an important platform to look at.  

Concretely, we look at comments on YouTube videos concerning the different candidates in the context of different search terms. First, we examine videos collected with general election search terms, such as #BTW21. Subsequently, we look at videos specifically focusing on the different candidates. 

Capturing the mood: Sentiment analysis  

Sentiment analysis acts as a first step to get a better idea of the overall mood of the discourse surrounding the elections. In order to capture sentiment, we use an expert-made sentiment dictionary to assign a value to each word in a comment. We then adjust for negation/emphasis, do some basic math, and come up with a score for the comment. 

Sentiment analysis: How it works 

The resulting scores can indicate negative feeling toward a video or its subject matter, which could be hateful speech, but often indicates negativity in general. For instance, videos about the floods in North Rhine Westphalia and Rhineland-Palatinate will have negative sentiment associated with them because of the tragic nature of the topic itself. Thus, we look at sentiment more as measuring the tone of the election and conversations around candidates. 

Average sentiment over time 

When looking at the numbers under the broadest hashtags (i.e., #BTW21), the mood seems relatively calm. For videos related to the election, the lion’s share of comments and words are neutral with negative and positive comments making up for a combined share of about one-third of all comments. The pattern is what statisticians call a normal distribution: some optimists, some pessimists, but most of us fall in the middle. 

Yet, this relative calm might be deceiving. In fact, when further specifying the search terms to include the different candidates running for the biggest parties, the outlook changes considerably. Despite there being fewer videos on candidates than the elections generally, these videos received twice as many comments and negative sentiment dominated. Personality is getting more attention than politics. 

Sentiment on the candidates

How should we interpret these negative comments? Sentiment scoring captures when people are talking about bad things (e.g., the state of the environment) or talk about the state of their party (e.g., think they are losing), not only hate speech. Based on this, supporters of Alice Weidel and Christian Lindner seem to have some sort of positive outlook. The people commenting on the videos of the other candidates, on the other hand, seem to have quite a negative view on either the world or their electoral chance (or both). 

Don’t be evil: Toxicity scores 

Finding hate speech online is difficult, especially in real-time as opposed to sifting through data after the election is over. Thus, for our ongoing analysis, we use Google’s Perspective API to classify comments. This tool - based on machine learning and used by many news organizations for content moderation - returns a probability that a comment is: toxic, severely toxic, an insult, an identity attack or a threat.  

Toxicity scoring: How it works 

As with our sentiment scores, Annalena Baerbock tops the ranks for toxicity and insults, as well as severe toxicity. However, Tino Chrupalla comes very close to Baerbock’s scores and surpasses her when it comes to attacks based on his identity and threats, although these are less common than toxicity and insults.  

Average toxicity over time 

Unsurprisingly, comments scored as highly toxic are quite harsh in their wording. Let’s look at what language is used.   

The bigger picture: Word associations 

While numbers are useful for getting a good grasp of the big picture, context matters when narrowing in on specific comments. Word associations allow us to get a deeper grasp of the narratives underwriting the varying perceptions.  

To find which words are used in reference to the candidates, we use a method called word embeddings (specifically Word2vec). This is a type of neural network, often classified as deep learning but with only two layers (so not so deep after all), which examines the context around the words in a text to derive relationships between these words. The end result can tell you about words that have similar meaning or are of a similar type - grouping for instance all the names of candidates together - or words that are associated with other words - like an adjective that is often used to describe a person.  

Word associations: How it works

Here again, we look at videos about #BTW21 generally and then focus on videos about the candidates from the most well-known parties. To get a more fine-grained result, we look not only at every word used, but also at the verbs, adjectives and nouns connected to the different candidates.  

Broadly, when looking at the words associated with the different candidates, the mood seems heated and polarised, reflecting what we discussed above. The comments on most candidates are associated with derogatory terms, insults, and negative attributes vis-à-vis the candidate’s competence and suitability to lead Germany, or lead anything at all. The varying characteristics attributed to the different candidates become especially pronounced when looking at the adjectives most associated with the different candidates.  

Personal attacks over content

Looking at the entire vocabulary of words used, the word associations more often show names and characteristics than content-related points. Even when there are content-related aspects, they mostly refer to scandals rather than policy proposals. For example, Olaf Scholz from the SPD is still mentioned in the context of the Cum-Ex and Wirecard finance scandals. Other political controversies such as his handling of the G20 summit as mayor of Hamburg in 2017 and his role in the Hartz IV labour market policy reforms do not (any longer) seem to play an important role. Similarly, for Annalena Baerbock from the Greens the most content-related points relate to her CV and some prominent gaffes during public speeches, rather than an in-depth contestation of her plans for a higher CO² price.  

All of these have in common that they are attention-grabbing issues, allowing us to quickly jump to conclusions (e.g. “Olaf Scholz in cahoots with a criminal and himself corrupt”, “Baerbock falsified her CV and plagiarised her book”). As such topics provoke the most reactions, they are likely to receive more engagement, get rewarded by the algorithms, and thus find their way to even bigger audiences. 

Thus, even if YouTube videos show the candidates talking about their manifestos and political agendas, a significant number of comments will contain expressions of opinions and attacks on a personal level rather than policy discussions. This is important to keep in mind, as it shows that political videos might be trending not due to the issue that is being discussed, but rather due to incendiary comments that encourage more people to engage with the video. Topics might thus be pushed to the forefront of attention and could cross social media platforms as trending ‘hot-topics’, even if the topic matter was not discussed by the online audience at all.  

Baerbock as a hated figure

The aforementioned personal attacks come to the fore when we turn to the videos collected by searching for the candidates and focussing on adjectives. The negative attributes used are quite similar across the board (e.g., associations close to incompetent, embarrassing, ridiculous, dumb). However, the traditionally dominant and current governing parties - the CDU and SPD - as well as the Greens receive the most heat. Once again, Annalena Baerbock is attacked in arguably the harshest manner.  

Examining the words associated with Baerbock more closely, besides the characteristics doubting her ability to be a chancellor, sexualised terms are used, belittling her as a female candidate. Interestingly, the same pattern cannot be observed with Janine Wissler of Die Linke or Alice Weidel of the AfD, two other leading female candidates. On the contrary, Alice Weidel’s attributes are by far the most positive of any candidate. The adjectives most commonly associated with her in the YouTube comments characterised her as intelligent, sincere, friendly, able, and candid. Only one other candidate comes close to this type of characterisation: Christian Lindner from the FDP.  

Der Tagesspiegel's special report on YouTube comments

Check out the article and cool interactive visual that our partners at Der Tagesspiegel have made using our shared data here!

Another perspective on the top words associated with the candidates

(Click on the picture to see it in 3D)

One can see that ‘ignorant’ is the number one adjective used to describe Tino Chrupalla, who amongst the candidates is the only one lacking a university education. Other candidates do not get this critique and instead are likely to be referred to as ‘incompetent’ or ‘arrogant’. 

As these results show, in addition to often succeeding at generating the most engagement, the AfD is also commented upon most favourably on YouTube. However, as seen with previous elections and the polls, this does not necessarily translate into electoral success.  

These results might therefore also tell us something about the different ways in which parties utilise YouTube to spread their ‘own news’. The videos on YouTube about the AfD candidates are especially successful at provoking positive engagement as many come from smaller channels catering to the AfD community. Conversely, many of the negative comments on other candidates can be found on videos by major news outlets. Candidates like Baerbock and Laschet have fewer niche channels cheering them on, even if they get wider coverage generally. 

Many press outlets do try to focus on the candidates, potential cabinet posts, and policies. But our social media habits and the platforms’ algorithms drive the news further towards personalities and candidates’ mishaps. As a result, most of the media coverage surrounding this year’s elections could focus on niche issues and scandals rather than the bigger picture and fundamental policy initiatives. Thus, it is more critical than ever for us to step out of our filter bubbles to be able to make informed decisions in the polls.  

This work is supported by

Stiftung Mercator