DLAB Seminar: Ordinal latent variables for learning intensity scales, by Niklas Stoehr

Concepts such as conflict intensity and sentiment polarity describe an intrinsic ordering, but are subjective and abstract. Thus, they may be best described using ordinal scales. They are however frequently treated as either nominal (categorical) or cardinal (continuous). We propose modelling intensity concepts using an ordinal latent variable model that can be learned from observable correlates. Particularly, we focus on conflict intensity: we learn an event-level intensity scale from observed perpetrator, victim and action types as well as fatality counts. Going beyond the level of individual events, we incorporate temporal dimensions of perceived intensity drawing connections to media attention and the notion of surprise.

Niklas Stoehr is a doctoral student at the Institute for Machine Learning at ETH Zurich advised by Ryan Cotterell and Bob West at EPFL. His interdisciplinary research aims at measuring latent intensity concepts. Combining methods from Natural Language Processing, Computational Social Science and Network Science, his interests are particularly centred around conflict intensity, numbers in text and sentiment analysis. He previously worked at the interface of these fields in industry (IBM AI Core, Microsoft Research, German Federal Foreign Office) and research (University College London, University of Oxford, Tsinghua University, TU Berlin).