SDS Workshop
The SDS Workshop is a space for SDS Hub participants to share work in progress, provide feedback on the work of other colleagues, and discuss project ideas. Papers for each workshop are circulated in advance for participants to review before each meeting.
Meetings are hybrid and Zoom links for each event are circulated on the SDS Hub mailing list prior to the event.
Past
2 November 2023 The Goldilocks Effect: The "Just Right" Writing Style of the Most-Used Global Corporate Responsibility Frameworks Adam Chalmers
The 21st century has seen an explosive increase in the number of global corporate responsibility (GCR) frameworks, issued by international organizations, that shape firms' communications of their commitment to corporate social responsibility (CSR). This study examines how firms' use of GCR frameworks may be affected by how GCR frameworks are written, by combining theoretical insights from affordance theory and computational linguistics, and using natural language processing methods to examine 320 firms' CSR communications (4,025 documents) for "re-use" of the GCR framework text. The study finds a "goldilocks effect" whereby firms' use of GCR frameworks is greatest when frameworks are written in language that is not too simple nor too complex.
Methods: NLP
10 March 2023 Machine Learning Models for the Measurement of Media Criticism Christopher Barrie
The ability of news media to criticize government is a core pillar of media freedom. Existing indices tend to use scoring criteria or expert surveys to develop over-time measures of media freedom. In this article, we use the largest existing dataset of Arabic-language news to evaluate how political reporting changes over the course of a successful and failed democratic transition. Using entirely unsupervised ALC word-embedding techniques, we demonstrate how to generate temporally granular measurements of media criticism that closely correlate with measurements derived from expert surveys for both Egypt and Tunisia. Crucially, the technique we propose is computationally inexpensive, effectively cost-free, and eminently scalable.
Methods: Word-embedding; NLP; Transformers