During the initial part of the project, the bulk of the work has gone into manual annotation of a series of articles on impacts of extreme events to be used as training and test data for the large language models. This was followed by the development of an automated pipeline for processing textual information. More specifically, the pipeline that we developed leverages large language models and in-context learning to extract semi-structured information from textual sources, which is normalized and refined in post-processing. A crucial step is geoparsing, which maps place names to geographical entities. An empirical evaluation based on the manually annotated benchmark data shows that extraction accuracy varies for different types of impact information, but is generally high.
Further work has focussed on connecting impacts of single instances of extreme events to the corresponding physical hazards. To do so, we first conducted a case-study analysis to understand how sensitive our results could be to the use of different metrics to identify the physical extremes. We have further developed a number of algorithms to automatically detect climate extremes, both in isolation and when multiple extremes co-occur. This work has evidenced the shortcomings of existing impact databases, whose impact entries often do not match hazards as recorded in historical climate data. We have specifically conducted an in-depth analysis of and intercomparison with the EM-DAT data. These results converged into a publicly available one-of-its-kind global database of impacts from multiple categories of climate extremes.
We have also performed work on weakly supervised language processing to derive indirect and cascading impacts. However, this has not yielded robust results with regards to the clustering of the impact categories and the ability of our pipeline to extract quantitative information on these impacts. We have therefore decided not to include these results in the initial version of the database. The scientific challenges encountered in this part of the work have resulted in a contribution to a perspective paper on the dynamics of multi-sector impacts of extremes. The inability to extract indirect and/or cascading impacts was a particularly high-risk high-gain part of the project, and had been identified as a potential implementation risk in the project plan. Consistently with this, the rest of the project was structured so as not to depend on this part of the analysis.
