A scalable and explainable approach to discriminating between human and artificially-generated text

Roberta Rocca, assistant professor, Interacting Minds Centre, Aarhus University

Abstract: With natural language generation models becoming increasingly fluent, being able to discriminate between human and artificially-generated text has become an urgent societal problem. However, existing approaches are inaccurate, non scalable, and uninterpretable, which makes them practically unusable in real-world contexts (e.g., detection of AI-generated essays) which require precision and accountability. We propose a novel and scalable approach to training text discrimination models based on interpretable linguistic and cognitive features. Using prompts from standard NLP benchmarks for paraphrase, dialogue, and summarization, we generate parallel corpora of human- and machine-generated text, train interpretable classifiers on linguistic and cognitive descriptors, and combine insights from resulting models and experimental evidence to highlight overlaps and differences in computational and human heuristics for text discrimination. 

Automatic neural machine translation for Greenlandic

Ross Deans Kristensen-McLachlan, associate professor, Department of Linguistics, Cognitive Science and Cognition, Aarhus University; Center for Humanities Computing, Aarhus University

Abstract: What is a word? This seems like a simple question but it continues to stump linguists and philosophers who spill millions of words trying to explain what they’re spilling. It seems, too, like it should be a concern for people who create language technology. How can teach computers to use words if we don’t even know what they are?

Contemporary natural language processing (NLP) makes assumptions about words which, by and large, are based on how major Indo-European languages behave. Those same linguists and philosophers might baulk but the engineer can reply with empirical results demonstrating the efficacy of their systems on goal-oriented language tasks. If it works, it works. But does it actually work?

This project tests these assumptions by applying modern language technology to a lesser-studied part of Denmark’s linguistic landscape – Greenlandic. This fascinating language of some 57,000 speakers exhibits many rare linguistic phenomena such as ergative alignment and polysynthetic morphology. Our goal is to train an automatic machine translation model for Greenlandic to Danish and back again. In doing so, we’ll empirically evaluate how well the assumptions of NLP hold up when applied to an extremely low-resource and morphologically complex language like Greenlandic.

Wool Worked Worlds – Studying Industrial Landscapes through Collaborative Filmmaking

Annika Capelán, Centre for Environmental Humanities and Department of Anthropology, Aarhus University

Abstract: A paradox characterises wool production: it is highly industrialised yet cannot be detached from place-specific ecologies and social relations. Wool sheared from sheep is materially shaped by the geopolitical and ecological particularities of the grasslands where they are raised. The EU Marie Cuire funded project Wool Worked Worlds explores how these local diversities shape the global wool industry, and how local sheep rearing and landscape management practices are shaped by practitioners’ imaginaries of geographically distant parts of the wool industry. Bringing together different versions of global wool, the approach deploys collaborative filmmaking engaging the work and analysis of sheep farmers, indigenous/black/coloured community members, laboratory technicians and ambulant sheep shearers from three wool regions – the Karoo and the Lesotho highlands of Southern Africa, Queensland, Australia and Chubut, Patagonia. The collaborative aspects of the film making implies that the work will be dialogical and that the consultants’ analyses will shape the film. IMC Seed funding adds a new layer of collaboration through the involvment of three junior scholars, all of them with attachments to the South African grasslands, and substantial research experience there. Through their skills and backgrounds, the coworkers add disciplinary diversity and allow for further methodological experimention which is key to the project questions about methods and analysis, i.e. about our abilities to respond to landscapes that are ‘wool worked’; damaged or otherwise affected by the longterm presence of woolen sheep along settler colonial dynamics. IMC Seed funding also stimulates stronger cross-institutional program building between AU CEH and UCT EHS.

Free of charge - All are welcome