Context and Moral Knowledge in Schwartz Value Detection

A recent investigation published on arXiv explores the influence of context and explicit moral understanding on the identification of Schwartz values in political texts. Researchers utilized the ValuesML/Touché ValueEval framework to analyze sentence, window, and full-document inputs, alongside no-RAG and retrieval-augmented configurations with a curated moral knowledge database. They employed supervised DeBERTa-v3-base/large encoders and zero-shot LLMs ranging from 12B to 123B parameters. Findings indicate that increased context does not always yield better results: while full-document context enhances supervised DeBERTa encoders by 3.8–4.8 macro-F1 points compared to sentence-only input, it does not consistently benefit zero-shot LLMs. Additionally, retrieved moral knowledge proves more effective in matched comparisons, enhancing all tested model families and context conditions under early fusion, though larger models show diminishing returns.