A systematic study investigates the role of context, external moral knowledge (retrieval), and model scaling in detecting Schwartz values within political texts. The research finds that context and retrieval are most beneficial for socially situated values, suggesting that context
Researchers conducted a systematic study on detecting Schwartz values in political texts, focusing on how context, explicit moral knowledge, and model architecture influence detection accuracy. The study compared various input methods (sentence, window, full-document), different knowledge integration techniques (no-RAG vs. retrieval-augmented), and model families ranging from DeBERTa to large LLMs (12B to 123B parameters).
The findings indicate that simply increasing context or model size does not guarantee performance gains across all setups. Specifically, full-document context significantly improved supervised encoders but offered no consistent benefit to zero-shot LLMs. Retrieved moral knowledge proved most consistently useful when used in matched comparisons. The study concludes that value-sensitive NLP systems should prioritize evaluating context, knowledge, and model family jointly rather than treating longer inputs or larger models as universal solutions. Furthermore, value-sensitive analysis showed that context and retrieval offered the greatest help for values that are socially situated or conceptually confusable.