Landslides often hit newsstands due to their destructive and potentially fatal effects. News are a valuable source of information for creating or enriching disaster databases and for expediting media-based studies of the dynamics of media attention. To accomplish that, news datasets must be filtered, geolocated and validated. This paper focuses on how landslides around the world are reported in German newspapers. We analyse almost 55k news articles about 4.5k news events in a 25-year period, compare it with external measures of countries’ susceptibility to landslides and provide insights, e.g. the overreporting of Southern and Western Europe, to foster further studies on inequalities in media attention to international disasters.

Climate change is acute in Central Asia (glacier loss, the Aral Sea, and drought exposure), yet there are no NLP resources for climate discourse in the region’s low-resource Turkic languages. We present CA-ClimateNews, the first human-annotated multilingual dataset for climate-relevance classification of Central Asian news, covering Kyrgyz, Kazakh, Uzbek, and (regionally grounded) Russian. The corpus comprises 3,140 news items labeled at two levels (binary climate relevance and an 8-way topic category) by two annotators per language, with a 100-item blind double-annotation overlap adjudicated to gold. Inter-annotator agreement is almost perfect (Cohen’s 𝜅 0.86–0.98 on relevance). Because we used transparent LLM-assisted pre-labeling, we additionally run a between-subjects experiment showing that displaying model pre-labels induces anchoring: an independent group shown the pre-label overrode it 3–7× less than a blind group. Our released labels are the blind version. We report baselines: fine-tuned multilingual encoders reach 88–96 relevance macro-F1 on the balanced benchmark; a Russian-only model loses up to ∼7 points transferring zero-shot to the Turkic languages even at matched training size (least, ∼2, for Kazakh), and a related Turkic source is no better, isolating a cross-lingual gap that in-language supervision closes. The dataset, guidelines, and code are publicly released.

Automated extraction of corporate greenhouse gas (GHG) emissions from sustainability reports is critical for climate crisis governance, yet the choice of prompt design, context provision, and model in LLM-based pipelines is largely guided by informal best-practice recommendations rather than systematic evaluation. With **Climatextract**, we assess these choices in a RAG pipeline, comparing closed-source against open-weight models, a structured "best-practice" prompt against a simple question-style prompt, and text-only against text-plus-table contexts. We find that performance is non-monotonic across model generations, a structured prompt with roles, definitions, extraction rules, and JSON formatting is significantly outperformed by a question-style prompt, and tabular context helps only in a specific model-prompt combination. A cost-benefit analysis shows that performance gains are often disproportionate to cost increases.

Historical archives are an under-used source for extending the instrumental climate record backward in time, and LLMs offer a way to extract the indices climatologists derive by hand. Beyond measuring how well systems extract this signal, we check whether their errors are safe to use for cross-century comparison, since a good correlation score does not rule out systematic, era-linked bias. Comparing lexical baselines, fine-tuned historical transformers, and LLM prompting on the Pfister temperature index across five centuries of German text, lexical methods beat every fine-tuned transformer we test, including one pretrained from scratch on historical German (r=-0.016). All six LLMs we test (Gemini 2.5 Flash, GPT-5-mini, DeepSeek v4 Flash, Claude Sonnet 4.6, Qwen3.7-Plus, Kimi-K2.6-Fast) show a warm bias that grows with calendar year, with the same sign in every model (slopes +0.13 to +0.34/century, p<0.01). The effect is modest in size (r²≈0.01–0.05) but consistent across six independently developed models. The best-correlated of the six, Gemini 2.5 Flash, matches the best lexical correlation (r=0.32) at double the error. An ablation stripping explicit dates and calendar-era markers from the quotes leaves this trend essentially unchanged, favoring an anachronistic present-day prior over the model correctly inferring the quote’s era. Correlation alone is thus insufficient for vetting an LLM as a historical-climate-index oracle.

As the environmental impact of Large Language Models (LLMs) grows, improving model efficiency has become critical for sustainable NLP. Recent advances in model compression have shown that ternary quantisation (i.e., using only {-1, 0, 1} as weights) substantially reduces memory footprint, inference time, and energy usage of LLMs with minimal performance degradation. However, whether these already efficient models can be further pruned to reduce resource requirements without compromising effectiveness remains unexplored. We investigate this question across six models, four sparsity levels, and six structured and unstructured pruning techniques. We find that even minimal pruning yields substantial efficiency gains: pruning fewer than 5% of parameters reduces inference time and energy consumption by up to 21%, and memory usage by up to 13%, with minimal accuracy degradation. We also find that, for pruned models with less than a 5% accuracy drop, the most energy and inference-time efficient configurations are achieved by pruning entire attention layers, whereas the most memory efficient configurations are obtained by pruning entire MLP layers. These findings highlight pruning as a complementary technique to ternary quantisation for reducing environmental footprint of LLMs while maintaining model effectiveness.

Plant species assemblages can contain climate-related information, but it is unclear whether plant language models remain useful when observations are incomplete. We compare a sparse LightGBM baseline with Pl@ntBERT and a tropically adapted variant combining vocabulary expansion and domain-adaptive pretraining. The models predict four WorldClim variables from species composition in São Paulo state under progressively shortened species lists. The clearest pattern occurs for precipitation of the driest month (BIO14). On the original split, A2-MT shows positive differences over LightGBM under severe incompleteness after hierarchical bootstrap and multiplicity correction. Because BIO14 was selected post hoc, we retrain the models on five alternative assignments of a nominal 40 km spatial-block design. LightGBM is higher on complete lists in four of five partitions, while positive BIO14 differences under incomplete lists are directionally more stable but retain corrected support in only two partitions. Additional checks show that the pattern persists with a longer context window, weakens under frequency-weighted removal, and narrows to Max-10 and Max-5 when both model families receive truncation-augmented training. These results show that within-split robustness does not substitute for retraining across alternative spatial partitions.

Slow-onset events such as droughts lack a clearly defined onset, and systematic databases do not capture their impacts at fine scales. News reports can help fill this gap by providing information on how drought-driven water shortages affect communities and how authorities respond. Yet, extracting structured information from text remains challenging. Here, we use news articles and official public communications on water-saving recommendations and water-use restrictions in Sweden as a case study to benchmark LLMs for drought-related information extraction. We compare the proprietary GPT-5 series against three large open-access models (Llama3.3, Deepseek, and GPT-OSS) in extracting variables such as the date, location, and causes of water-saving measures. Using prompt engineering, structured outputs, and minimal post-processing, we evaluate performance using a test set of 210 manually annotated articles. The GPT-5 model series outperforms all open-access models across all variables. We find that, without task-specific fine-tuning, open-source models are sometimes outperformed by a keyword-based baseline, suggesting that general-purpose capabilities alone are insufficient for reliably extracting structured data. These findings underscore the challenges of monitoring water-shortage risks based on texts, particularly when extracting complex variables.

Auditable analysis of corporate climate disclosures requires retrieving, for each regulatory disclosure requirement, the report page or section that contains the evidence. We study ontology-aware query expansion (OAE) for ESRS E1 (Climate Change) requirements over a new page/section-level benchmark: five companies, 1,177 real-artifact chunks, and 114 human-reviewed gold labels, plus a held-out ESRS-native company (64 labels) unseen during ontology construction. Against lexical, dense, hybrid, and pseudo-relevance-feedback (PRF) baselines, a lightweight ontology built from the standard’s requirements improves retrieval: the best configuration reaches Recall@10 0.842 vs. BM25’s 0.702 (McNemar p=.003). Generic PRF expansion hurts retrieval while ontology-guided expansion helps; a corpus-blind LLM query expander is a far stronger control and ties with the ontology in-domain, but does not transfer to the held-out company, where OAE stays ahead of lexical, dense, and LLM-expansion baselines while being statistically indistinguishable from cross-encoder reranking — a comparison we treat as diagnostic rather than conclusive. We provide an error analysis of the residual misses. "Gold" denotes retrieval ground truth only; we make no company scoring or compliance judgement.

Contracts serve as a cornerstone mechanism of private environmental governance, and yet their form as unstructured text has meant a recent overhaul in their construction and iteration due to large language models (LLMs). As such, historical forms of knowledge sharing in the legal field must also change. We present the Climate Contracting Knowledge Graph, a publicly released resource that organises the model clauses, glossary terms, and guides of The Chancery Lane Project (TCLP) into a queryable property graph, linked to the laws and standards they invoke. Existing climate knowledge graphs target policy documents or climate-model metadata; none represent the contractual instruments through which private climate commitments are implemented. The graph is built with a lawyer-in-the-loop pipeline and achieves a statistically significant improvement in mean average precision over dense and lexical baselines. The graph, models, and a lightweight retrieval service are released publicly.

Sustainability reports denote greenhouse gas emissions with heterogeneous unit strings, such as abbreviations, inconsistent formatting, and scale markers. In emission extraction pipelines, these units must be normalized to a common target, e.g. t CO2e, to enable reliable comparison and aggregation of emissions across companies, reporting years, scopes, and regions. We propose a hybrid normalization approach: If a raw unit string is not found in a curated dictionary, a machine learning classifier maps noisy unit strings to normalized emission-unit categories that can then be converted to the target unit. The classifier is trained using character-level TF-IDF, lightweight string features, and LLM-assisted data augmentation. On random-splits of the augmented dataset, an ensemble model achieves approximately 95% accuracy. Additional evaluation on rare real-world unit strings indicates that supervised fallback classifiers can handle out-of-dictionary variants in climate-related information extraction.

Climate change is reshaping food systems and consumer practices, creating new challenges for food safety. Understanding how people modify their eating habits in response to climate disruptions and sustainability challenges is critical for assessing emerging microbiological risks. We present FoodClimNER, a new named entity recognition (NER) corpus containing extracts from open-access scientific articles on food science in the context of climate change, with manual annotations made by an expert. We define annotation guidelines with expert definitions of food practice entities and climate change entities. We evaluate baselines and a custom prompting method with recent LLMs on this task in a zero-shot setting. Results indicate that expert-level entity types and the presence of nested and discontinuous mentions make this a challenging testbed, with significant room for improving zero-shot methods. We also provide a detailed analysis of the importance of different parts of our custom prompt.

News media play a central role in shaping public perceptions of climate change, and whether coverage emphasizes threats or solutions affects audience engagement and policy support. We compare fine-tuned encoders and zero-shot prompting for classifying sentences from German-language climate news as threatoriented, solution-oriented, both, or neither, using two independent binary classifiers. We finetune three German encoders with context windows ranging from the target sentence alone to four preceding and four following sentences, and compare them with three open-weights generative models and the non-generative decision model Jev. All models are evaluated with article-grouped five-fold cross-validation on 440 manually coded Austrian newspaper articles. The best encoder, GBERT-large with the widest context, achieves weighted F1 scores of 0.805 (threat) and 0.802 (solution). The recently released MiMo-V2.6-Flash reaches 0.807 and 0.798 without task-specific training, placing both approaches on par, whereas earlier prompting models remain below the best encoder. Additional context improves all encoders consistently, by up to 4.8 F1 points. On the natural class distribution, MiMo-V2.6- Flash outperforms GBERT-large on threat classification, while GBERT-large retains a small advantage for solution classification. These results indicate that the gap between fine-tuned encoders and prompted language models has only recently been closed.

Monitoring climate change adaptation (CCA) in the heritage sector, especially in the context of World Heritage (WH) management, requires systematic analyses of policy documents for CCA strategies, a process that is difficult to scale. We ask whether automated retrieval systems can prioritise expert annotation effort: given a CCA concept, can a system rank passages of a management plan so that those an expert would annotate appear near the top, without any in-domain supervision? This objective is formulated as an information retrieval problem with a new benchmark dataset on climate-adaptive heritage management strategies, created from expert-labelled text chunks in official Management Plans of 11 Dutch WH properties. Four retrieval approaches (BM25, a bi-encoder, a dense retriever with reranker, and a cross-encoder reranker) are evaluated and compared. The cross-encoder reranker achieves the strongest performance across labels. An LLM-as-judge evaluation further shows that top-ranked unlabelled chunks could be relevant, suggesting that the existing annotations are precise but not exhaustive. Per-label analysis identifies which concepts are automatically retrievable and which remain beyond the reach of general-purpose retrieval. Our benchmark establishes a foundation for scalable mapping and monitoring of climate change adaptation in the context of heritage management.

Corporate sustainability reports include a content index mapping each disclosure required by a reporting framework (e.g. GRI) to its supporting pages — a resource that potentially benefits nearly every downstream sustainability-NLP task, yet no existing benchmark scores its recovery as a framework-grounded, page-resolved output. We formalize Content Index Detection (CID) as three evaluable stages — index page retrieval (S1), table extraction and disclosure linking (S2), and page-label-to-index resolution (S3) — and release esg_cid_plus: 37 GRI reports and 7,294 rows stratified across three difficulty tiers, from unchanged real reports to an adversarial tier of naturally hard documents and synthetic perturbations. Our main benchmark sweeps open-weight vision-language models from sub-1B to 35B parameters (plus frontier-API references) as S2 extractors under shared, deterministic S1 and S3 stages: the best backends reach S2-F1 of 0.89–0.92 on the easier tiers, while standard OCR-benchmark scores explain only a third of the variance in extraction quality. Composing extraction with the deterministic resolver exposes S3 as the open frontier: under label-scheme resolution alone, where even best extraction caps end-to-end F1 at 0.11 on the hardest tier. esg_cid_plus will be released in huggingface.

Difference-in-differences (DID) studies are widely used to evaluate climate policy, but assessing the evidence supporting their identification assumptions remains challenging. We introduce ARGUS, a structured language-model pipeline that audits reported evidence against an eleven-dimension assumption–implication–evidence rubric and abstains when relevant evidence cannot be retrieved. We evaluate ARGUS using injected flaws, economics papers, and a small pilot with reconciled labels. On the 11-flaw benchmark, ARGUS detects 73% of planted flaws, compared with 18% for a keyword-based pipeline. Across 26 economics papers, ARGUS abstains on about 40% of paper–dimension assessments for lack of retrievable evidence. In a five-paper pilot with labels reconciled by two annotators, it assigns a higher risk level than the labels on 25 of the 33 assessments it completes. A rule fixed before the labels arrived removes most of this in-sample; weighted agreement stays low. ARGUS provides evidence-linked risk reports that localize potential weaknesses for expert review, without adjudicating causal claims. Code and data: https://github.com/yonghongzhang-io/ARGUS

As large language models (LLMs) are increasingly used, understanding their energy footprint has become an important concern. The growing adoption of retrieval-augmented (RAG) based pipelines for climate-domain specific analysis raises a key question: how does the energy consumption of domain-specific RAG-based pipelines compare with that of direct generic LLM usage? Prior research has focused on standalone model calls or coarse token-based estimates, while leaving the energy implications of deployed application workflows insufficiently understood. In this paper, we assess the inference-time energy consumption of two LLM-based climate analysis chatbots (ChatNetZero and ChatNDC) compared to the generic GPT-4o-mini model. We estimate energy use under actual user queries by decomposing each workflow into retrieval, generation, and hallucination-checking components. We also test across different times of day and geographic access locations. Our results show that the energy consumption of domain-specific RAG-based pipelines depends strongly on their design. More agentic pipelines substantially increase inference-time energy use, particularly when used for additional accuracy or verification checks, although they may not yield proportional gains in response quality. This study provides a new understanding on how the design of domain-specific LLM products affects both the energy footprint and quality of output.

Information seeking systems have yet to access multimodal information about thematic maps in scientific literature. Thematic maps are images of geographical maps showing themes of information and are crucial for understanding the impact of climate on natural resources. Although Multimodal Large Language Models (MLLMs) are applied in different applications, their effectiveness in identifying thematic maps is not explored. We address this gap by evaluating some of the state-of-the-art MLLMs to (1) classify thematic maps using different visual and textual contexts; (2) explain their predictions and uncover their sources of bias using _attribute discovery_; and (3) assist humans with the annotation of thematic maps in the scientific literature. We release a new public dataset (https://github.com/tomtx/cp-thematic-maps) and use it to show that MLLMs that were investigated in our study have room for improvement. We observe that MLLMs can be biased toward visual attributes of thematic maps (graphical indicators). The best classification result is an 80% strict F1 score by the Grok 4.3 model, which can fluctuate between 76% and 84% due to classification bias. Overall, we consistently find that the input of images with captions and examples of the correct answers yields substantial gains in task performance.

Water utilities communicate both routine service information and time-sensitive public-risk guidance, yet multilingual NLP benchmarks for institution-authored infrastructure communication remain limited. We introduce AquaCrisis, a corpus of 20,000 text-bearing Facebook posts from 12 European water utilities, with a 1,500-post manually annotated gold standard. We formulate Task A as classification of a post’s main water-utility topic and Task B as classification of its primary communicative function, together with a coarser grouped Task B.1. We compare rule-based and supervised classifiers with four instruction-tuned LLMs on native-language and English-translated inputs. Supervised approaches achieve the highest point estimates on the fine-grained tasks: translated TF-IDF with logistic regression performs best on Task A, while translated multilingual sentence embeddings perform best on Task B. GPT-4o-mini achieves the highest point estimate on grouped Task B.1. Paired bootstrap analysis shows that translation effects are model- and setting-dependent: gains are clearest for the lexical TF-IDF classifier and selected Gemma 4 and GPT-4.1-mini conditions, while multilingual sentence embeddings, GPT-4o-mini, and Qwen 3.5 are comparatively stable across input representations. AquaCrisis provides a benchmark for multilingual water-utility communication spanning routine, educational, operational, and incident-related messaging, supporting communication monitoring, operational triage, and cross-lingual evaluation.

We present a German-language corpus for studying divisiveness in political social-media discourse: 7,277 parent–child comment pairs sampled from the YouTube channels of two ideologically opposed German parties, the AfD and Die Linke, on two contested policy domains, climate change and migration. Each pair is annotated on three dimensions: interaction quality (a five-class constructive–destructive ×agreement–disagreement scale), stance (of parent and child toward the topic), and the presence of 14 propaganda techniques. We detail the full construction pipeline– channel selection, comment and transcript collection, pseudonymization, conversation-tree reconstruction, topic filtering, and stratified sampling– and a reproducible three-task annotation protocol: a 700-pair human gold standard validates an LLM-assisted pipeline that scales annotation to the full corpus. Descriptive analysis reveals pronounced party- and topic-specific patterns– most notably that the Migration × AfD stratum is the most hostile (54.5%ofits pairs are Destructive Agreement) while thread-level quality does not differ significantly between the two topics (χ2(3) = 5.15, p =0.16). Only on climate are exchanges between opposing sides (a skeptic replying to a supporter) constructive on average; the same clash on migration remains destructive, and on both topics the most destructive exchanges occur between two opponents of the policy at issue.

Corporate sustainability reports communicate environmental disclosures through text, tables, charts, and infographics, making assessment against sustainability reporting standards challenging. Existing disclosure extraction systems focus on information extraction but do not determine whether disclosed evidence satisfies reporting requirements. We study requirement-level multimodal environmental disclosure alignment, a task that evaluates whether sustainability report evidence satisfies GRI 305 and ESRS E1-4 and E1-6 requirements. To support this task, we curate an expert-annotated benchmark comprising 731 disclosure-alignment instances from 43 sustainability reports. Using this benchmark, we evaluate three Vision Language Models (VLMs): GPT-5.5, Gemini, and LLaMA Maverick under three conditioning strategies: open-ended, framework-specific, and knowledge-augmented. Results show that although frontier VLMs possess substantial implicit knowledge of sustainability reporting standards, they frequently produce over generalized and optimistic alignment judgments. Requirement level grounding improves alignment accuracy by 12% and yields more accurate disclosure gap identification. Our findings demonstrate the value of knowledge grounded multimodal reasoning for scalable, interpretable, and auditable sustainability disclosure assessment.

Tracking progress on climate change adaptation requires systematic and robust annotation of adaptation policies, but manual annotation has become infeasible due to the rapidly growing volume of unstructured documents. Pre-trained language models (LMs) have emerged as promising tools for automating and scaling annotation. However, relatively little is known about their accuracy and quality on complex, free-text annotation tasks. Meanwhile, large LMs (LLMs) are increasingly deployed for large scale text analyses despite their considerable environmental footprint and limited exploration of more sustainable alternatives. By benchmarking model accuracy against human inter-annotator agreement, reviewing perceived annotation quality, and estimating energy and water consumption, we show that fine-tuning small LMs (SLMs) provides a substantially more sustainable alternative to LLMs with only marginal reductions in accuracy and improved perceived annotation quality. Furthermore, we demonstrate the importance of evaluating complex annotation tasks beyond accuracy against human gold standard annotation as ambiguity and human error can obscure meaningful differences in model performance. By reducing the costs of high-quality policy annotation, our approach enables the scalable knowledge extraction needed for tracking progress on adaptation, living evidence syntheses, and evidence-informed policy evaluation.