Grounding Political Discourse in the Body: A Computational Approach to Analyzing U.S. Energy Narratives

Image schemas, recurring pre-conceptual gestalts arising from sensorimotor experience, constitute the foundational level of embodied conceptual architecture, yet their systematic identification in naturalistic language data remains a persistent methodological challenge. Traditional introspective analyses in cognitive linguistics yield rich theoretical insights but face difficulties in scalability, inter-coder reliability, and systematic replication. This study addresses these challenges by developing a computationally assisted analytical framework that integrates semantic role labeling (AllenNLP) with large language model verification (DeepSeek R1) to detect image-schematic activations in discourse. Grounded in Fairclough’s three-dimensional critical discourse analysis model, the pipeline extracts narrative triples (“agent–action–patient”) from 10 U.S. energy policy documents (74,426 characters, 289 sentences), clusters them into conceptual domains via SentenceBERT embeddings and DBSCAN, and subjects the outputs to human–LLM collaborative validation. Against a manually annotated gold standard (50 triples from 29 sentences), the pipeline achieves an F1 score of 0.79, a 20-percentage-point improvement over the raw SRL baseline. Three image schemas (CONTAINER, PATH–GOAL, and FORCE) are identified as systematically activated through specific discourse strategies. By providing structured SRL output as input to the LLM rather than requesting annotation from scratch, the pipeline reduces hallucination while preserving the directional semantic relations needed for schematic analysis. This study offers a replicable, scalable methodology for investigating embodied language in naturally occurring discourse.