A natural language processing framework for mining paleoclimate literature and its application for global synthesis of 2000-year climate change reconstructions

Despite extensive efforts in paleoclimate reconstruction, globally integrated analyses of historical climate changes over the past two millennia remain constrained by the inefficiency of traditional manual data collection and synthesis methods. To address this limitation, we propose a scalable natural language processing (NLP) framework leveraging pre-trained language models (PLMs) to systematically retrieve, annotate, and extract structured information from paleoclimate literature. We first constructed a corpus of 770,821 abstracts collected from ResearchGate, filtered using strict Boolean keyword matching. A subset of 850 abstracts (including negative samples to mitigate background noise) was manually annotated to train a BERT-BiLSTM-CRF named entity recognition (NER) model. This automated pipeline successfully extracted 3657 climate-anomaly entities together with their associated locations, temporal boundaries, and semantic categories. To synthesize these decentralized textual records, we developed the Temperature Voting Index (TVI) and Precipitation Voting Index (PVI) to quantify the large-scale temporal consensus of temperature and hydroclimate anomalies. Comparative analysis showed varying degrees of agreement between the TVI and the PAGES 2 k temperature reconstructions over the full Common Era ( r = 0.24–0.75), although the weakest correlation was not statistically significant. Agreement was substantially stronger during 1500–2000 CE, when all correlation coefficients were 0.80 or higher, providing an external consistency check on the text-derived temperature index. Furthermore, the PVI demonstrated high sensitivity to historical hydroclimatic extremes, particularly severe droughts. While exploratory comparisons revealed multidecadal phase alignments between the PVI and the North Atlantic Oscillation (NAO), we emphasize that such text-derived indices predominantly reflect historical documentation intensity and inherent reporting and extraction biases, rather than definitive shifts in atmospheric circulation dynamics. Our findings highlight the potential of advanced NLP methods to transform paleoclimate data synthesis by enabling rapid, large-scale knowledge extraction from unstructured text. The proposed framework establishes a reproducible and scalable text-mining foundation for interdisciplinary geoscience research.