Automated extraction of scientific statements for integrated assessment of animal welfare using language models

• We propose an NLP framework that supports semantic modelling of animal-welfare literature. • Our end-to-end Transformer models extract scientific welfare statements. • We employ adversarial contrastive sampling to mitigate negative-sample scarcity. • On benchmarks, E5 achieves 96.0% precision and 92.4% recall. • We provide an intuitive interface for interactive exploration of extracted statements. Animal welfare is a key concern in livestock farming, and a semantic modeling approach for welfare assessment is recognized as scientifically grounded. However, extracting supporting scientific evidence still relies heavily on manual review of the literature, limiting scalability and efficiency. To address this challenge, we propose a natural language processing (NLP)-based framework for the automatic extraction and classification of welfare-related scientific statements. To mitigate the lack of negative samples, we adopt a positive-unlabeled (PU) learning strategy enhanced by adversarial contrastive sampling, enabling the automatic extraction of high-confidence negative samples from unstructured documents. Three classification paradigms are evaluated: (1) few-shot learning with large language models (LLMs), (2) a two-stage method combining Transformer-based encoders with traditional classifiers, and (3) end-to-end fine-tuning of pre-trained Transformer models, including BERT, RoBERTa, BGE, and E5. Our experiments reveal that the end-to-end models achieve the best performance, with the E5 reaching 96.0% precision and 92.4% recall on the test set. The adversarial sampling strategy also demonstrates strong robustness across different similarity thresholds. In a challenging scenario where semantic differences between positive and negative statements were small, the BGE-based model outperformed the other models, achieving 59.1% precision and 76.5% recall. To support practical deployment, we also introduce an interactive visualization interface for real-time statement extraction and inspection. Overall, the proposed framework offers a scalable and generalizable solution for extracting welfare-related scientific statements, thereby advancing AI-assisted modeling in animal welfare research. The source code is available at https://github.com/jay-zc/SM_classification.git (as of September 2025).