News
GIST @ EMNLP 2026: Presenting Three Papers
29 Oct 2026
New work on making LLM-based extraction of climate data more reliable.
GIST (Greenhouse Gas Insights and Sustainability Tracking) research project is presenting three papers at the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP), one of the leading conferences in natural language processing: one at the Industry Track and two at the ClimateNLP Workshop. All three build on our work on extracting greenhouse gas emissions from corporate sustainability reports, a project we carry out in cooperation with the Deutsche Bundesbank.
How Should We Design Prompts for Information Extraction?
Anna Steinberg Schulten, Malte Schierholz, L. Reichenbach, Frauke Kreuter, and Alexander Fraser present “A Practitioner’s Guide to Systematic Prompt Evaluation for Information Extraction” at the Industry Track. Using greenhouse gas emissions extraction as a case study, the paper tests prompt design choices such as structure, context, and model in a systematic way. One finding is that structured JSON prompts can underperform simpler question-style prompts. The results are condensed into a decision guide for practitioners.
Climatextract: Putting Prompts to the Test on Our Benchmark
At the ClimateNLP Workshop, the same team presents “Climatextract: A Systematic Prompt Evaluation for Climate Information Extraction”. Where the Industry Track paper distills guidance for practitioners, this paper shows the evaluation at work: prompt variants are compared on our benchmark dataset for greenhouse gas emission extraction, published in Scientific Data, to see how prompt design affects the quality of the data extracted from sustainability reports.
Making Emission Units Comparable
Also at the ClimateNLP Workshop, L. Reichenbach, S. Oehler, Malte Schierholz, and Anna Steinberg Schulten present “Beyond Dictionaries: Machine-Learned Normalization of Emission Units in Information Extraction”. Emissions are reported in many different units and spellings, which makes extracted values hard to compare. The paper replaces hand-built dictionaries with machine-learned normalization.