All projects
2026

CARE-RA: Reasoning-Augmented Argument Mining for Renewable Energy Acceptance

A neuro-symbolic argument mining framework that explains why communities accept or oppose renewable energy projects, combining structured argument extraction with chain-of-thought reasoning and faithfulness verification.

Overview

When communities oppose wind farms, solar installations, or other renewable energy infrastructure, the reasons matter as much as the outcome. Understanding whether resistance stems from concerns about land rights, visual impact, noise, or perceived procedural unfairness has direct implications for how developers, policymakers, and energy justice advocates should respond. Most NLP approaches to this problem stop at sentiment classification — flagging support or opposition — without explaining the underlying reasoning.

CARE-RA (Community Acceptance of Renewable Energy — Reasoning-Augmented) goes further by integrating explicit reasoning chains into the argument mining pipeline. The goal is not just to detect a stance, but to explain it in a way that is faithful to the source evidence and grounded in energy justice theory.

Technical Approach

The project makes contributions at three levels: dataset, model architecture, and evaluation framework.

CARE-RA Dataset: A new benchmark combining approximately 300 expert-annotated gold articles with ~3,000 LLM-annotated silver articles. Each annotation captures structured argument components — the community subject (C), energy project (E), specific concern (CC), stance, and impact — along with explicit chain-of-thought justifications and energy justice frame labels (Distributional, Procedural, and Recognition justice categories).

RA-AM Architecture: The Reasoning-Aware Argument Mining model consists of four integrated components:

  • Structured Argument Extractor: Identifies the key argument elements (C, E, CC) and predicts stance and impact relations
  • Dual-Stream Reasoning Encoder: Processes verbal/rhetorical cues separately from factual/logical content, capturing both dimensions of how arguments are made
  • Chain-of-Thought Reasoning Generator (CoT-ER): Produces explicit reasoning traces that connect evidence spans to the predicted stance
  • Faithfulness-by-Unlearning (FUR) Verification Module: Tests whether each reasoning step causally influences the model’s prediction, filtering out post-hoc rationalizations

Evaluation Framework: Beyond standard F1 scores for argument structure accuracy, the project introduces multi-dimensional faithfulness evaluation: FUR scores (percentage of causally significant reasoning steps), ROSCOE metrics for semantic and logical quality, and counterfactual robustness tests.

Significance

This work addresses one of the core criticisms of neural NLP systems applied to social science problems: that they produce predictions without intelligible justifications. The FUR verification module is particularly novel — it operationalizes faithfulness not as a property of the generated text but as a causal property of the reasoning process itself.

The immediate application is an Explainable Social Acceptance Index (ESAI) for renewable energy projects — a structured, evidence-grounded measure of community sentiment that policy analysts and project developers can actually use to understand and address local concerns. The methodology extends naturally to other domains where understanding the reasoning behind public attitudes matters as much as measuring those attitudes.

This project builds on the earlier CARE framework (2025) and draws on climate risk analysis work developed through the GDELT climate risk project (2024).