More Context, Larger Models, or Moral Knowledge? A Systematic Study of Schwartz Value Detection in Political Texts

A systematic study of Schwartz value detection in political texts finds that adding context and explicit moral knowledge can improve performance, but larger language models do not consistently deliver better results, according to research submitted for review in May 2026 [1]. The work, led by Víctor Yeste, uses the ValuesML/Touché ValueEval format to compare sentence-level, window-based, and full-document inputs across supervised DeBERTa-v3 encoders and zero-shot large language models ranging from 12 billion to 123 billion parameters [1][2]. The study shows that full-document context improves supervised DeBERTa encoders by 3.8 to 4.8 macro-F1 points over sentence-only input, but the same additional context does not consistently help zero-shot LLMs [2][3]. Retrieved moral knowledge proved more uniformly beneficial. Under an early-fusion retrieval-augmented generation setup, adding curated moral-knowledge snippets improved every tested model family and context condition [2][4]. Gemma, Qwen, and Mistral all improved with retrieval for sentence, window, and document prompts, even when longer context alone had degraded no-RAG performance [3][4]. Paired bootstrap intervals over test sentences remained above zero for all LLM RAG contrasts and for all DeBERTa-v3-base RAG contrasts [3]. Scaling model size did not guarantee gains. Moving from DeBERTa-v3-base to DeBERTa-v3-large and from 12-billion-parameter LLMs to much larger variants did not automatically lift scores under the tested protocol [1][2]. Simple early fusion also outperformed the late-fusion and cross-attention RAG variants evaluated for encoders [2][4]. A model card on Hugging Face for the DeBERTa-v3-base document-plus-early-RAG checkpoint trained with seed 1701 reports a macro-F1 of 0.3224 and micro-F1 of 0.3617 on the held-out test split [5]. Across three seeds the same condition averaged a macro-F1 of 0.314 with a standard deviation of 0.008, which the paper identifies as the strongest aggregate result among the tested supervised encoders and zero-shot LLMs [5]. The moral knowledge base used for retrieval contains short, paraphrased chunks based on Schwartz value definitions, annotation guidance, and theory-level contrasts [5]. Per-value analyses indicate that context and retrieval help most for socially situated or conceptually confusable values [1][2]. The authors conclude that value-sensitive NLP systems should evaluate context, knowledge, and model family jointly rather than treating longer inputs or larger models as universal improvements [2][4].