Answers are grounded in this paper and cite their sources.
Pilot study demonstrates cognitive search enhancements in political discourse, suggesting new retrieval methods.
IS-RAG: Image-Schematic Retrieval-Augmented Generation for Cognitive Search in Political Discourse This deposit contains a pilot study (proof of concept) investigating whether explicit indexing of image schemas (Lakoff & Johnson, 1980; Talmy, 1988) — pre-linguistic patterns such as FORCE, CONTAINER, and PATH — can bridge the lexical-cognitive gap in dense retrieval: cases where query and document express the same conceptual structure through different vocabularies. During indexing, an LLM-based Cognitive Parser annotates each chunk with the image schemas present, stored as structured metadata alongside dual vector embeddings (text and cognitive). At query time, schemas detected in the query trigger a proportional boost over chunks sharing the same cognitive structure. The evaluation uses a TREC-style ground truth over 121 Brazilian parliamentary speeches (403 chunks, 9 valid queries, NDCG@5), with a fully deterministic pipeline (temperature = 0) and an independent LLM annotator. Main findings. Text-mode boosting is the configuration closest to the baseline (NDCG@5 0.7490 vs. 0.8012, Δ = −0.052); using cognitive embeddings as search vectors severely degrades retrieval (Δ = −0.502), localising the technique's value in metadata-based re-ranking rather than dual-embedding search. Three methodological findings are reported: (1) schema frequency in the corpus does not predict boostability; (2) queries must be formulated as assertive metaphorical statements to activate the cognitive component; (3) the gain is conditional on the base embedding quality — stronger models (BAAI/bge-m3) close the lexical-cognitive gap implicitly, defining a utility frontier for the technique in resource-constrained or locally-hosted deployments. The study is presented as a negative-result-informative pilot that opens a research line in cognitive information retrieval, releasing the methodological scaffolding (cognitive metadata indexing, proportional boosting, schema-stratified TREC-style evaluation) and, to our knowledge, the first Portuguese parliamentary corpus annotated with image schemas and graded relevance judgements. Files: main.pdf (English) and main-ptbr.pdf (Portuguese, Brazil). Code and data: github.com/yanjustino/is-rag Versão em português: este depósito contém um estudo piloto que investiga se a indexação explícita de esquemas imagéticos (FORCE, CONTAINER, PATH) pode superar a lacuna léxico-cognitiva na recuperação densa, avaliado sobre 121 discursos parlamentares brasileiros. Os resultados localizam o valor da técnica no re-ranqueamento por metadados cognitivos e caracterizam sua fronteira de utilidade frente a embeddings mais fortes.
No takes yet. Share an insight, caveat, or question.
Yan de Lima Justino (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: