Pinned Repositories
CV
ifkw_mmDL
Recent developments in the machine learning field has enabled reserachers to incorporate audio and visual materials into their analysis by using deep learning. These methods are especially valuable as multi-modal political communication has increasingly become more ubiquitous with the advent of web 3.0 and similar technologies. There is, however, very little empirical work evaluating how applicable and beneficial these methods are in automated content analysis in political communication research. This study aims to fill this gap by comparing classification performances of shallow learning and deep learning algorithms using multi-modal embeddings and tabular data. To test their performance, we classify 898 tweets from EU executives into binary category using a set of deep learning and shallow learning algorithms. In our experiments, shallow learners with tabular data have regularly outperformed the deep learning classifier using multi-modal embedding. Our results tell a few cautionary tales about using multi-modal representation for future researchers. First, most visual political communication have large variation. Therefore, multi-modal automated content analysis requires large amount of manually annotated data. Compounded with the general data-gready nature of deep learning, it can be more reseource efficient to use shallow learners with monomodal tabular data for classification tasks. It is also possible to use multi-modal embeddings with shallow learners such logistic regression and support vector machine. However, these algorithms are not capable of handling high dimensional datasets such as tensors (i.e embeddings). With an appropriate dimension reduction method applied to a tensor, these might yield the optimum results yet there is more research needed to identify such dimension reduction methods
Paper1_scripts
Paper2_communication_quality
Given the politicization of European integration, effective public communication by the European Union (EU) has gained importance. Especially for rather detached supranational executives, social media platforms offer unique opportunities to communicate to and engage with European citizens. Yet, do supranational actors exploit this potential? This article provides a bird’s eye view by quantitatively describing almost one million tweets from 113 supranational EU accounts in the 2009–2021 period, focusing especially on the comprehensibility and publicity of supranational messages. We benchmark these characteristics against large samples of tweets from national executives, other regional organizations, and random Twitter users. We show that the volume of supranational Twitter has been increasing, that it relies strongly on the multimedia features of the platform, and outperforms communication from and engagement with other political executives on many dimensions. However, we also find a highly technocratic language in supranational messages, skewed user engagement metrics, and high levels of variation across institutional and individual actors and their messages. We discuss these findings in light of the legitimacy and public accountability challenges that supranational EU actors face and derive recommendations for future research on supranational social media messages.
Paper3_scripts
Scripts used for data collection, cleaning and analysis for the third paper of my dissertation. Scripts are not yet ordered, contact the author via sina.ozdemir@ntnu.no if you would like to run them yourself
Paper4_engagement
Scripts used to generate and analyze data for the fourth paper of my dissertation
Phd_data_collection
This collection of scripts were used to collect data from Twitter for my phd project
shinySurveys
Code to produce a variety of R Shiny surveys
SICSS2021HELSINKI
This repo contains scripts for programming tasks in SICSS Helsinki 2021
SOS3510
SinaOzdemir's Repositories
SinaOzdemir/ifkw_mmDL
Recent developments in the machine learning field has enabled reserachers to incorporate audio and visual materials into their analysis by using deep learning. These methods are especially valuable as multi-modal political communication has increasingly become more ubiquitous with the advent of web 3.0 and similar technologies. There is, however, very little empirical work evaluating how applicable and beneficial these methods are in automated content analysis in political communication research. This study aims to fill this gap by comparing classification performances of shallow learning and deep learning algorithms using multi-modal embeddings and tabular data. To test their performance, we classify 898 tweets from EU executives into binary category using a set of deep learning and shallow learning algorithms. In our experiments, shallow learners with tabular data have regularly outperformed the deep learning classifier using multi-modal embedding. Our results tell a few cautionary tales about using multi-modal representation for future researchers. First, most visual political communication have large variation. Therefore, multi-modal automated content analysis requires large amount of manually annotated data. Compounded with the general data-gready nature of deep learning, it can be more reseource efficient to use shallow learners with monomodal tabular data for classification tasks. It is also possible to use multi-modal embeddings with shallow learners such logistic regression and support vector machine. However, these algorithms are not capable of handling high dimensional datasets such as tensors (i.e embeddings). With an appropriate dimension reduction method applied to a tensor, these might yield the optimum results yet there is more research needed to identify such dimension reduction methods
SinaOzdemir/CV
SinaOzdemir/Paper1_scripts
SinaOzdemir/Paper2_communication_quality
Given the politicization of European integration, effective public communication by the European Union (EU) has gained importance. Especially for rather detached supranational executives, social media platforms offer unique opportunities to communicate to and engage with European citizens. Yet, do supranational actors exploit this potential? This article provides a bird’s eye view by quantitatively describing almost one million tweets from 113 supranational EU accounts in the 2009–2021 period, focusing especially on the comprehensibility and publicity of supranational messages. We benchmark these characteristics against large samples of tweets from national executives, other regional organizations, and random Twitter users. We show that the volume of supranational Twitter has been increasing, that it relies strongly on the multimedia features of the platform, and outperforms communication from and engagement with other political executives on many dimensions. However, we also find a highly technocratic language in supranational messages, skewed user engagement metrics, and high levels of variation across institutional and individual actors and their messages. We discuss these findings in light of the legitimacy and public accountability challenges that supranational EU actors face and derive recommendations for future research on supranational social media messages.
SinaOzdemir/Paper3_scripts
Scripts used for data collection, cleaning and analysis for the third paper of my dissertation. Scripts are not yet ordered, contact the author via sina.ozdemir@ntnu.no if you would like to run them yourself
SinaOzdemir/Paper4_engagement
Scripts used to generate and analyze data for the fourth paper of my dissertation
SinaOzdemir/Phd_data_collection
This collection of scripts were used to collect data from Twitter for my phd project
SinaOzdemir/shinySurveys
Code to produce a variety of R Shiny surveys
SinaOzdemir/SICSS2021HELSINKI
This repo contains scripts for programming tasks in SICSS Helsinki 2021
SinaOzdemir/SOS3510
SinaOzdemir/TBMM_discourse
Democracy in the Mediterranean EU candidate countries has considerably deteriorated in recent years. This article draws on the Turkish case to assess how democratic backsliding affects parliamentary EU membership discourses. In response to the governing AKP‘s authoritarian turn, we attest increasing support for EU membership among opposition parties that are not allied with the incumbent. We also find that the predominant framing of Turkey‘s EU membership in the Turkish Parliament shifted from normative to utilitarian arguments, emphasizing cost-benefit considerations instead of liberal democratic values. Since parties‘ ideological predispositions remained relatively stable over the period under study, we pinpoint governmentopposition dynamics as the primary mechanism explaining changing party positions and frames on EU membership. The article concludes with an outlook and a discussion of implications beyond the Turkish case.