Principled Analysis of Energy Discourse across Domains with Thesaurus-based Automatic Topic Labeling

Principled Analysis of Energy Discourse across Domains with Thesaurus-based Automatic Topic Labeling Thomas Scelsi1 Alfonso Mart´ınez Arranz2 Lea Frermann1 1 School of Computing and Information Systems, The University of Melbourne 2 Department of Chemical Engineering, The University of Melbourne tscelsi@student.unimelb.edu.au {alfonso.arranz,lfrermann}@unimelb.edu.au Abstract With the increasing impact of Natural Lan- guage Processing tools like topic models in social science research, the experimental rigor and comparability of models and datasets has come under scrutiny. Especially when con- tributing to research on topics with worldwide impacts like energy policy, objective analyses and reliable datasets are necessary. We con- tribute toward this goal in two ways: first, we release two diachronic corpora covering 23 years of energy discussions in the U.S. En- ergy Information Administration. Secondly, we propose a simple method for automatic topic labelling drawing on domain knowledge via political thesauri. We empirically evalu- ate the quality of our labels, and apply our la- belling to topics induced by diachronic topic models on our energy corpora, and present a detailed analysis. 1 Introduction Policy-making in highly technical areas such as energy is deemed to require neutral input from spe- cialised government agencies, e.g. the US’s Energy Information Administration (EIA) or the Interna- tional Energy Agency (IEA). Their publications are known as“grey literature”, given that they are ex- pertly produced but not peer-reviewed. In contrast to academic literature, grey literature is often freely available online and aims to be more accessible to lay readers. Energy grey literature has been shown to display biases towards incumbent fossil fuel technologies (Mart´ınez Arranz, 2016; Mohn, 2020), but no thor- ough exploration exists of the disconnect between grey and academic literature in their coverage of energy issues. In this paper, we provide a repro- ducible and automated assessment of topics across both literatures. We analyse and make available two diachronic corpora derived covering 23 years of energy discus- sions from two EIA publications: the International Energy Outlook (IEO) and the more nationally fo- cused Annual Energy Outlook (AEO). Parsing re- ports of government agencies is non-trivial due to their diverse layouts (changing over time), and text frequently disrupted with tables and graphs. We release a clean text corpus as a basis for future re- search on energy communication. We also make available a software tool-set for the creation and analysis of these corpora, easily generalisable to diachronic datasets in general. We analyse the corpora using dynamic topic models (Blei and Lafferty, 2006). We compare their discussion of various facets of energy politics against the discourse in the scientific community over the same period, drawing on a corpus of ab- stracts of articles published in established energy journals. By releasing our grey literature corpus, we ad- dress the reproducibility challenge of large-scale text analyses with unsupervised models like topic models in the social sciences (M¨uller-Hansen et al., 2020). Studies are often difficult to reproduce, com- pare against or build upon due to a lack of public resources, as well as ad-hoc subjective choices of the researchers. We address the …

Content truncated to 500 words