Training a language model for low-resource languages is challenging due to data scarcity and computational cost. Tokenizer transfer offers a way to adapt a pre-trained model to a new tokenizer without full retraining, improving efficiency and cross-lingual applicability. To the best of our knowledge, we present the first controlled evaluation of tokenizer transfer on monolingually pretrained base models trained on language-specific corpora. We evaluate Orthogonal Mapping Pursuit (OMP) and Fast Vocabulary Transfer (FVT) across six languages and multiple fine-tuning regimes. We computed byte-normalized log-perplexity and MultiBlimp accuracy for the Goldfish model family. We use these to evaluate for target-language adaptability and source-language retention. We add monolingual or mixed finetuning to compare with only using transfer. OMP with monolingual target finetuning achieves the best target-language performance, yielding lower log-perplexity and higher MultiBlimp scores than all evaluated baselines. These include (i) a model trained only on the source language, (ii) a model trained on a smaller amount of target-language data, and (iii) the source-language model adapted via standard finetuning on the target data. The results suggest tokenizer transfer is a compute-efficient alternative for low-resource LM training: train a monolingual tokenizer for the target language, transfer it to a larger pre-trained model, and fine-tune using the target data. Code: github.com/skogsgren/tokeneval

Large language models (LLMs) have demonstrated remarkable performance across a wide range of NLP tasks. However, their effectiveness in discourse parsing remains underexplored where the existing LLM-based attempts fall significantly short of the performance achieved by the encoder-based models. In this study, we propose a Chain-of-Thought (CoT) prompting approach for the task of implicit discourse relation recognition (IDRR), leveraging the concept of abstract objects. We show that guiding the model to identify abstract objects within the arguments of the discourse relation systematically enhances the classification performance across both Level-1 and Level-2 senses, in both monolingual and multilingual settings. Through experiments on three monolingual and one multilingual corpora, covering seven languages and annotated according to PDTB 3.0, we demonstrate that our CoT-style prompting approach achieves significant improvements over previous LLM-based methods.

Climate extremes like storms, heatwaves, wildfires, droughts and floods significantly threaten society and ecosystems. However, comprehensive data on the socio-economic impacts of climate extremes remains limited. Here we present Wikimpacts 1.0, a global climate impact database built by extracting information from Wikipedia using natural language processing. Our method identifies relevant articles, extracts the information using GPT4o, post-processes the information and consolidates the database. Impact data is stored at the event, national, and sub-national levels, covering 2726 events from 1034 to 2024, with 17 912 national and 32 343 sub-national entries. The database shows low error scores (range from 0 to 1) for event-level information like timing (0.05), deaths (0.03), and economic damage (0.12), and slightly higher error scores for injuries (0.21), homelessness (0.25), displacement (0.29), and damaged buildings (0.28) compared to manually annotated data from 156 events. Wikimpacts 1.0 provides a different event coverage than EM-DAT, notably providing broader coverage of storm impacts but more limited coverage of flood impacts. We match 179 events between the two databases to compare impact values, and find that 32 out of 179 matched events have identical data for deaths, and 7 out of 77 for injuries. However, there are notable discrepancies in information on homelessness and damage. We view the publicly available Wikimpacts 1.0 database as a complementary resource to existing impact databases, which facilitates subnational climate impact assessments, and highlights the potential of natural language processing to enhance existing impact datasets and provide robust information on climate impacts. Lastly, we provide a static version of the database as used in this paper at https://bolin.su.se/data/li-2025-wikimpacts-1.0.final (last access: 20 May 2026) and the Wikimpacts website (https://www.wikimpacts.eu/, last access: 20 May 2026) for visualization and access for future updates of the database

Book reviews are instrumental in assessing the public value of literary works. However, the path to literary recognition, considering the factors influencing why certain books are reviewed, earn acclaim, or face criticism, remains unclear. Amid ongoing discussions around the “crisis of criticism,” the study scrutinizes this discourse and employs two regression analyses to examine the occurrence and sentiment of reviews for 9814 fiction books originally in Swedish and 8340 reviews from major Swedish newspapers (2001–2018). The sentiment of book reviews is determined through a novel approach with an automated pipeline incorporating state-of-the-art natural language processing models. The article reveals enduring institutional, cultural, and demographic hierarchies that shape literary recognition through the occurrence of reviews. Key factors include literary prestige and author youth, alongside ties to elite literary networks, publication by major houses, and focus on literary fiction or poetry. Conversely, older authors with numerous books, bestsellers, or a focus on children's literature exhibit reduced review probabilities. While review occurrence is highly predictable, sentiment proves less so. Moreover, reviews generally express positivity despite frequent reviewer discord. Our findings suggest that books are selected for review based on external factors such as reputation and credentials, thereby reflecting persistent patterns of cultural consecration. This resonates with the idea of a “conservative revolution”, where established hierarchies maintain their influence, rather than indicating a “crisis in criticism”. Nevertheless, the unpredictability of review sentiment and the lack of consistent consensus on quality underscore a deeper evaluative uncertainty that transcends the more stable hierarchies governing review selection

Chemosensory sensations are often hard to describe and quantify. Language models may facilitate a systematic understanding of sensory descriptions. We accessed consumer and expert reviews of wine, perfume, and food products (English language; about 68 million words in total) and analyzed their sensory descriptions. Using a novel data-driven method based on natural language data, we compared the three chemosensory vocabularies (wine, perfume, food) with respect to their vocabulary overlap and semantic properties, and explored their semantic spaces. The three vocabularies primarily differ with respect to domain specificity, concreteness, descriptor type preference and degree of gustatory vs. olfactory association. Wine vocabulary primarily distinguishes between white wine and red wine flavors and qualities. Food vocabulary separates drinkable and edible food products and ingredients, on the one hand, and savory and non-savory products, on the other. A salient distinction in all three vocabularies is between concrete and abstract/evaluative terms. Valence also plays a role in the semantic spaces of all three vocabularies, but valence is less prominent here than in general olfactory vocabulary. Our method allows a systematic comparison of sensory descriptors in the three product domains and provides a data-driven approach to derive sensory lexicons that can be applied by sensory scientists. 

ClimateEval is a comprehensive benchmark designed to evaluate natural language processing models across a broad range of tasks related to climate change.ClimateEval aggregates existing datasets along with a newly developed news classification dataset, created specifically for this release.This results in a benchmark of 25 tasks based on 13 datasets, covering key aspects of climate discourse, including text classification, question answering, and information extraction.Our benchmark provides a standardized evaluation suite for systematically assessing the performance of large language models (LLMs) on these tasks.Additionally, we conduct an extensive evaluation of open-source LLMs (ranging from 2B to 70B parameters) in both zero-shot and few-shot settings, analyzing their strengths and limitations in the domain of climate change.

Large Language Models (LLMs) have demonstrated an impressive ability to retrieve and summarize complex information, but their reliability in conflicting contexts remains poorly understood. We introduce an adversarial extension of the Needle-in-a-Haystack framework in which three mutually exclusive “needles” are embedded within long documents. By systematically manipulating factors such as position, repetition, layout, and domain relevance, we evaluate how LLMs handle contradictions. We find that models almost always fail to signal uncertainty and instead confidently select a single answer, exhibiting strong and consistent biases toward repetition, recency, and particular surface forms. We further analyze whether these patterns persist across model families and sizes, and we evaluate both probability-based and generation-based retrieval. Our framework highlights critical limitations in the robustness of current LLMs-including commercial systems-to contradiction. These limitations reveal potential shortcomings in RAG systems' ability to handle noisy or manipulated inputs and exposes risks for deployment in high-stakes applications

Extreme weather events increasingly challenge the reliability and resilience of modern power systems. This study presents a novel methodology for assessing weather-related disruptions by integrating two large-scale datasets: outage statistics from Swedish energy companies and meteorological hazard data from the Swedish Meteorological and Hydrological Institute (SMHI). A key contribution is the use of large language models to extract and structure relevant information from unstructured SMHI weather reports, creating a structured hazard database. The datasets are merged and analyzed in Power BI, with a focus on wind-related events. The analysis identifies over 123 000 outages linked to wind and wind-induced treefall, offering valuable insights into spatial and temporal patterns of disruption. A multi-day event window is introduced to improve event-outage matching, and the IEEE Major Event Day (MED) methodology is applied to classify extreme disruption days. The results contribute to a deeper understanding of the impact of weather on power systems and lay the groundwork for future research

Evaluation of Grammatical Error Correction (GEC) systems is becoming increasingly challenging as the quality of such systems increases and traditional automatic metrics fail to adequately capture such nuances as fluency versus minimal edits, alternative valid corrections compared to the 'ground truth', and the difference between corrections that are useful in a language learning scenario versus those preferred by native readers.Previous work has suggested using human post-editing of GEC system outputs, but this is very labor-intensive.We investigate the use of Large Language Models (LLMs) as post-editors of English and Swedish texts, and perform a meta-analysis of a range of different evaluation setups using a set of recent GEC systems.We find that for the two languages studied in our work, automatic evaluation based on post-editing agrees well with both human post-editing and direct human rating of GEC systems.Furthermore, we find that a simple n-gram overlap metric is sufficient to measure post-editing distance, and that including human references when prompting the LLMs generally does not improve agreement with human ratings.The resulting evaluation metric is reference-free and requires no language-specific training or additional resources beyond an LLM capable of handling the given language.

Whereas human cognition develops through perceptually driven interactions with the environment, language models (LMs) are “disembodied learners” which might limit their usefulness as model systems. We evaluate the ability of LMs to recover sensory information from natural language, addressing a significant gap in cognitive science research literature. Our investigation is carried out through the sense of smell — olfaction — because it is severely underrepresented in natural language and thus poses a unique challenge for linguistic and cognitive modeling. By systematically evaluating the ability of three generations of LMs, including static word embedding models (Word2Vec, FastText), encoder-based models (BERT), and the decoder-based large LMs (LLMs; GPT-4o, Llama 3.1 among others), under nearly 200 training configurations, we investigate their proficiency in acquiring information to approximate human odor perception from textual data. As benchmarks for the performance of the LMs, we use three diverse experimental odor datasets including odor similarity ratings, imagined similarities of odor pairings from word labels, and odor-to-label ratings. The results reveal the possibility for LMs to accurately represent olfactory information, and describe the conditions under which this possibility is realized. Static, simpler models perform best in capturing odor-perceptual similarities under certain training configurations, while GPT-4o excels in simulating olfactory-semantic relationships, as suggested by its superior performance on datasets where the collected odor similarities are derived from word-based assessments. Our findings show that natural language encodes latent information regarding human olfactory information that is retrievable through text-based LMs to varying degrees. Our research shows promise for LMs to be useful tools in investigating the long debated relation between symbolic representations and perceptual experience in cognitive science.