Fig 2. Results of the latent Dirichlet allocation, with thematic dominance and clustering visualized on the left, and the baseline frequency of keywords across the entire corpus depicted on the right. This was generated using the freely available pyLDAvis software in the GENSIM Python package.
3.3 LDA topic quadrants
We labeled the axes of the principal component quadrants to define overarching context for the scenarios, to ensure that topics which are close to one another in the quadrant space are similar in some way, while those far apart are dissimilar. To be clear, this step is entirely user-defined, and relies on human interpretation of the text analysis. For the purposes of our scenarios, which aim to explore thefuture of the Arctic, we use two broad features that would serve as exogenous drivers for the entire region: climate change severity and regional levels of cooperation.
The x-axis describes the relative severity of climate change in the Arctic region, with the right side representing anticipated climate changes, and the left side representing extreme climate changes. Given the overwhelming evidence to date of how the Arctic is disproportionately sensitive to climate change (Dodds, 2010; Post et al., 2009; Steiner et al., 2019; Stroeve et al., 2012), the concentration of most of the scenarios in that domain is supported empirically. The left side of the x-axis represents more extreme climate changes, which we acknowledge are possible in the range of climate projections, as well as being exceptionally consequential to the region.
The y-axis describes whether the Arctic region is characterized by conflict or cooperation. The top of this axis represents high cooperation, including economically, politically, socially, and militarily. The bottom of this axis represents low cooperation. In this way, the bulk of the thematic clusters are somewhat neutral, and this is broadly reflected in the patchy history of Arctic cooperation (Osherenko & Young, 2005). There have been periods of peace and international cooperation, exemplified, for example, by the effectiveness of the Arctic Council (Young, 2010). Likewise, there have been profound periods of conflict (Keil, 2014; Rahbek-Clemmensen, 2017; Young, 2011), as well as targeted violence toward Indigenous peoples (Crawford, 2014; Salusky et al., 2021).
To be clear — the position of the topic clusters in this quadrant space does not determine the content of the scenario but helps to frame the context of the scenario. Thus, when the scenarios produced in this analysis are compared with one another, those that were identified computationally as more tightly clustered exist in similar types of geopolitical and climatic contexts (i.e., topics 1, 2, 3, 4, 5, 6), whereas those that were distributed throughout the intertopic space will be more idiosyncratic (i.e., topics 7, 8, 9, 10).
3.4 Distribution of topic clusters in the Intertopic Distance Map
We can clearly interpret the current dominance of each topic relative to the others, with topics 1-6 being quite dominant and topics 7-10 being less so. This information provides additional context of how representative a given theme is for the entire corpus. Here we provide a summary showing each topic with its corresponding keywords (Table 1).