digicomlab Master Thesis Funding Grants


Since 2018 the digicomlab has a reoccurring call for applications for the Digital Communication Methods Lab Master Thesis Grant. These grants provide financial support for theoretically-relevant and digitally innovative (research) master theses written at the University of Amsterdam’s Graduate School of Communication.

This pages provides a list of currently funded theses and completed theses that received funding in 2020 and 2019.


Funded theses in progress (2021)

Affective and Behavioral Effect of Explaining Recommendation Contrastively in Chatbot: Role of Anthropomorphism by Zhimiao Li


The significance of chatbots for product recommendation as a novel communication tool for marketing purposes has been widely recognized by businesses across a variety of verticals.  However, existing recommender systems are seriously deprived of transparency into the recommendation process which has been increasingly inherently complicated and uninterpretable where outputs are made in a “black box”. This may result in consumer skepticism towards the recommendations from a chatbot. 

Accordingly, accompanying recommendations with contrastive self-explanation as a state-of-art approach to explainable artificial intelligence (XAI) has been proposed to ease the problem and raise user’s acceptability of the recommendations. Yet, no research on the extent to which providing a contrastive explanation influences on the affective and behavioral outcomes of user interaction with a product recommendation chatbot has been conducted. 

On the theoretical basis of the elaboration likelihood model (ELM) and the information adoption model (IAM), the current study explores the affective and behavioral effect of contrastive explanation in the context of chatbot. Moreover, the study investigates the moderating role of anthropomorphic cues. By employing an online experiment and a chatbot developed with the conversational agent research toolkit (CART), the results of the study are expected to contribute to the line of research in the persuasiveness of chatbot in recommendation services.


Understanding users’ responses to communication with disclosed vs. undisclosed customer service chatbots: A mixed methods study by Nathalie Koubayová


With the recent rising of chatbots in e-commerce, the conversational agents entered our lives, without many people noticing. However, because of huge advancements in the humanisation of AI, it is now difficult for people to assess whether it is a chatbot or a human that they communicate with. That brings ethical and privacy issues, as it is not obligatory for companies to disclose the identity of their chatbots. Because disclosure at the start of the conversation negatively impacts purchase figures due to consumers’ perception of conversational agents being less knowledgeable and emphatic as opposed to human agents, many businesses rather keep chatbots’ identity a secret. As there is a likelihood that the disclosure of a chatbot’s identity will be enacted by law in the European Union, inspired by the California Consumer Protection Act, this thesis aims to reach a better understanding of the impacts of chatbot disclosure, using a mixed method approach. 

This will be achieved by conducting qualitative interviews and a quantitative experiment, in which respondents will be exposed to a disclosed or undisclosed version of a chatbot, created with Conversational Agent Research Toolkit (CART), to investigate the impact on three types of users’ responses: anthropomorphism, social presence, and source orientation. 

 By investigating these concepts, this thesis hereby contributes in-depth insights into limited existing research about the effects of chatbots’ disclosure. These insights are pivotal for societal stakeholders as knowledge of how disclosures foster and inhibit users’ responses to chatbots is needed for the ethical design of contemporary chatbots.   


The Consequences Of Conspiracies: Isolating The Effect Of Social Media Conspiracy Exposure On Political Participation For Self-selecting Individuals by Timothy Dörr


With the rise of social media (SM), conspiracies, while not a new phenomenon, are spreading faster and farther than ever before. Consequences of such conspiracies are also becoming more apparent, for example through the anti-corona demonstration or the US January 6th insurrection. Despite this, research into conspiracy consequences is still sparse, and the research that does exist only looks at overall population effects. 

This is problematic since previous research has demonstrated that estimating overall population effects for stimuli prone to self-selection obtain results that can hardly be generalized outside of the lab setting. When investigating the effects of social media conspiracy exposure, self-selection is prominent as this high-choice media environment allows ‘selectors’ to choose to see conspiracies, and ‘non-selectors’ to avoid them. Therefore, my research will attempt to investigate the effect of conspiracy theories on government trust, voting intention, and protesting intention, while also decomposing these treatment effects into selector and non-selector effects. 

I will do so by employing a survey experiment that mimics a social media feed. However, instead of randomly dividing subjects into treatment and control conditions only (where they are forced to watch the stimulus unless they exit the experiment), my design also employs a third group where participants are allowed to self-select into treatment or control. This third group will give me information on the proportion of the population that does self-select into conspiracies, which helps me to decompose the treatment effect for selectors and non-selectors. 


Polarized news, polarized audiences? A quantitative content analysis of digital political news and user comments in India by Apeksha Shetty


Polarization has come to characterize politics in several countries around the world. While a certain degree of polarization may indeed contribute to electoral stability and help consolidate new democracies (Lupu, 2015), if this phenomenon reaches a relatively high degree of intensity, it can lead to the corrosion of democratic systems and contribute to public disaffection with political parties (Carothers & O’Donohue, 2019). In India, the rise of partisan media in traditional and digital environments is believed to contribute to polarization between groups with a vision of a Hindu nation and groups with a more secular view. While some scholars have commented on polarization in their research on Indian social media (Neyazi, 2020; Udupa, 2019), the links between polarization in news media and the impact on their audiences remain understudied, particularly on Facebook. The analysis of such social media data offers the unique opportunity to observe audience reactions to news posts as they are published and consumed.

This study aims to deploy natural language processing tools developed in the last few decades, from Wordfish to Spacy to BERT, to the problem of discovering and characterizing polarized text written by journalists and audiences of media outlets across the left-right political spectrum, covering multiple news events. To the best of the applicant’s knowledge, this is the first study of its kind to explore polarization on Indian social media, and Facebook in particular, from different perspectives.


A social network analysis approach to identifying topic communities and influentials in German Twitter discourse concerning Covid-19 vaccinations by David Leimstädtner


During the course of the Covid-19 pandemic, Germany has seen the rise of the anti-vaccination movement. But how has this influx of new supporters unfolded in the digital space? To answer this, my thesis examines the differences in network structures and information flow between vaccine advocating and vaccine hesitant topic communities on Twitter and how these structures changed throughout the pandemic. To achieve this, a number of network models are calculated based on a dataset containing all German-language tweets containing keywords of the vaccination discourse. The new Full Archive Search of the Twitter API 2.0 is thereby used to collect all Tweets since the beginning of the pandemic.

Using the resulting networks, a community detection algorithm is run in order to identify separate clusters of users within the vaccination discourse. Based on existing literature, these are hypothesized to include a vaccination-hesitant community as well as a large majority of vaccination advocates. To identify the positions of the resulting clusters, a machine-learning classifier is trained to concern between pro and anti vaccination content. Finally, the clusters compared in terms of structural differences in information flows and content-related differences (such as most influential users and commonly posted sources). 


What’s in a voice? The effects of vocal attractiveness and vocal dominance on the outcome of actual elections by Lonneke van Riele

In 2016, Hillary Clinton, nicknamed “Shrillary”, was criticized for having a too high-pitched voice and Margret Thatcher infamously lowered her voice pitch during her time in office. Not just the contents of what politicians communicate, but also how they communicate has an effect on voting behaviour (Banai, Banai, & Bovan, 2017; Banai, Laustsen, Banai, & Bovan, 2018). Previous research has shown that winners of actual elections have, on average, lower voices and less pitch variability than their opponents (Banai et al.,2017; Banai et al., 2018). However, although voice pitch data resulting from automated voice analysis can provide us with a lot of interesting information, it cannot provide us with an explanations of these characteristics are preferred by voters. On the one hand, it has been argued that humans tend to choose leaders that show signs of physical dominance (Laustsen & Petersen, 2015). On the other hand, it has been argued that lower pitched voices may be preferred by voters because they are found to be more attractive (Leaderbrand), leading to more favourable evaluations of these politicians (Surawski, & Ossoff, 2006).

To gain more in-depth knowledge about the reason why people prefer certain vocal characteristics in their leaders, a mix of automated voice analyses using Praat (phonetic voice analysis programme) and quantitative methods will be used. As vocal attractiveness cannot be determined in an automated way using PRAAT, it will be coded manually. Furthermore, the influence of voice pitch and pitch variability will be examined in an experiment.


The effect of humour in political messaging: An investigation combining fine-tuned neural language models and social network analysis by Janice Butler

Humour is a key factor for influencers and can be a powerful initiator of information propagation. In a commercial sense, it promotes consumer engagement and in a political setting the right kind of humour evokes a positive, “man of the people” effect. Current research on natural language understanding is based on the use of large data sets and relies on deep learning. There has been some research on humour in language based on word embeddings but very little using neural language models – techniques which have only been developed in the last 3 years and whose rapidly improving performance on benchmark tasks are, even today, not entirely understood. Although researchers have been able to provide binary determination of humour (funny/not funny), the measurement of type (fun, nonsense, wit, irony, satire, sarcasm, cynicism) and degree of humour has hardly been tackled. In the project, fine-tuning of natural language models such as BERT, XLNet and GPT-2 aims to provide a tool to do just this.

Since the intention is to identify effective strategies in political messaging with the aid of humour, the final phase of the project correlates types and degrees of humour with success in terms of effectiveness in social media. Conducting a network analysis of large numbers of political tweets, determining which tweets attract the most activity as measured by the level of their betweenness centrality facilitates a mapping to the level and type of humour determined by the tooling developed in the first phase of the project for comedic detection.


News Recommenders and Democracy: Assessing the User Experience of Democratic News Recommender Systems  by Nordin Bouchrit

News recommenders are increasingly depicting the news consumption of the audience by thoroughly selecting and directing the most fitting information to the audience. It is hypothesized that news recommenders can potentially offer opportunities for the democratic role of the media. On the other hand, the so-called filter bubbles caused by the isolation of the user in a restricted information environment is a potential anti-democratic outcome of a news recommender. 

In essence, news recommender systems are neither good nor bad. The outcome is heavily dependent on the design of the system and the values upon which they are developed. Although previous studies show that recommendation systems not necessarily lead to a reduction in diversity, the assessments that are being made are mostly evaluative in nature by comparing different recommendation approaches without real user-interaction. Moreover, previous studies that are done with a real user-interaction are mostly of a commercial nature optimizing the recommender system in order to increase the revenue of the recommendation executer. Therefore, effects on users when being confronted with a recommender system that is designed based on improving the diversity to benefit democratic values remain understudied.

The aim of this thesis is to fill this gap by designing news recommenders that try to advance values and goals essential in a democratic society and test the effects on the user-experience in a realistic setting by using the 3bij3 framework.



Completed Funded Theses


Acquiring Political Knowledge through Meme Exposure on Facebook: An Eye Tracking Experiment by Julia Dalibor

Do Individual Differences Moderate the Effect of Political Attacks Conveyed via a Deepfake on Evaluations of Politicians? by Mónika Simon

Viral Violence: The Effects of Police Violence Framing, Group Identity, and Militarization on Public Outrage and Perception of Police  by Neil Fasching

Designing Virtual Reality Experiences to Promote Pro-Environmental Behaviour: The Longitudinal Effects of Prompts by Hana Hegyiova


Online vigilance and goal conflict stress smartphone users out: An in-situ approach to digital stress by Alicia Gilbert

Assessing political news quality: An automated comparison of political news quality indicators across German newspapers with different modalities and reach by Nicolas Mattis

Emotions as the Impetus of Negative Campaigning Effects. Investigating the mediating role of anger and fear in the effects of negative campaigning by Vladislav Petkevic



Can AI trust you? Cultivating perceptions of trustworthiness in online Conversational Agents by Eirine Ntaligkari

Disclosing or Disguising Influencer Marketing on Instagram? by Céline Müller

Fake News in the 2018’s Brazilian Presidential Election: An analysis of its diffusion on Twitter by Pieter Attema Zalis

Mapping the Issue Arena of Plastic Soup: Applying the Network Agenda-Setting (NAS) Model in Big-Data Research by Louelle Pesurnaij

What’s really the difference? Developing machine learning classifiers for identifying Russian state-funded news in Serbia by Ognjan Denkovski


Consuming Less Meat For Me or For Us? by Demi van der Plas

Expectation Hurts? by Xiaotong Chu

Immersive Persuasion by Jonas Schlicht

The Sound of Immersion by Noëlle S. Lebernegg