Enhancing Language Models for Robust Greenwashing Detection

Sustainability reports are critical for ESG assessment, yet greenwashing and vague claims often undermine their reliability. Existing NLP models lack robustness to these practices, typically relying on surface-level patterns that generalize poorly. We propose a parameter-efficient framework that structures LLM latent spaces by combining contrastive learning with an ordinal ranking objective to capture graded distinctions between concrete actions and ambiguous claims. Our approach incorporates gated feature modulation to filter disclosure noise and utilizes MetaGradNorm to stabilize multi-objective optimization. Experiments in cross-category settings demonstrate superior robustness over standard baselines while revealing a trade-off between representational rigidity and generalization.

💡 Research Summary

The paper addresses the pressing problem of detecting greenwashing and vague claims in corporate sustainability reports, a task that is increasingly important for reliable ESG assessment. Existing natural‑language‑processing (NLP) approaches for ESG analysis tend to rely on shallow lexical or stylistic cues, which work well on in‑distribution data but fail when companies deliberately rephrase or obscure their disclosures. To overcome this limitation, the authors propose a parameter‑efficient, structured representation learning framework that is built on top of low‑rank adapters (LoRA) and integrates three complementary components: (1) supervised contrastive learning that pulls together semantically related ESG claims while pushing apart unrelated ones; (2) an ordinal ranking loss that encodes the three‑level actionability continuum (indeterminate → planning → implemented) defined in the A3CG benchmark, ensuring that higher‑actionability claims are embedded closer to each other than lower‑actionability ones; and (3) a per‑sample gating mechanism that dynamically balances the contribution of the contrastive and ordinal objectives for each training instance.

The contrastive loss uses cosine similarity with a temperature hyper‑parameter τ and supports multiple positive examples per anchor, which is well‑suited to the noisy, multi‑aspect nature of ESG texts. The ordinal loss adopts a mean‑margin formulation with a fixed margin m₀, encouraging a minimum distance gap between embeddings of different action levels. To prevent either loss from dominating, the gating module computes scaled scores for each loss (using temperatures T_ctr and T_ord), applies a softmax, and yields normalized weights w_ctr(i) and w_ord(i) for every sample i. This fine‑grained weighting allows the model to emphasize the most informative signal for each claim.

Training both losses simultaneously creates a classic multi‑objective optimization problem where gradient magnitudes can become imbalanced. The authors therefore introduce MetaGradNorm, an extension of GradNorm that dynamically rescales loss weights based on the relative difficulty of each task. They compute the gradient norm for each loss, define a difficulty ratio r_k from the current loss relative to its initial value, and set target gradient norms G*_k = \bar{G}·r_k^γ, where γ controls sensitivity. A meta‑objective J(α|θ) penalizes deviations between actual and target gradient norms and includes an entropy regularizer (weighted by β) to keep the gating distribution from collapsing. Both model parameters θ and meta‑parameters α (including λ_base, λ_ord, T_ctr, T_ord, γ, β) are updated jointly via stochastic gradient descent, with a softplus reparameterization to enforce positivity.

The entire training pipeline consists of two stages. In stage one, the LoRA adapters are trained with the combined contrastive, ordinal, gating, and MetaGradNorm components, shaping the latent space to reflect graded actionability. In stage two, the same adapters are fine‑tuned on the downstream ESG claim classification task without re‑initialization, preserving the structured representation while adapting to the specific supervision signal.

Experiments are conducted on the Aspect‑Action Cross‑Category Generalization (A3CG) dataset, which pairs sustainability aspects with action descriptions labeled as indeterminate, planning, or implemented. The authors adopt a cross‑category evaluation protocol: certain aspect categories are held out during training to form an “unseen” test set, while the remaining categories constitute a “seen” test set. Six open‑source LLMs are evaluated: T5, LLaMA‑3‑8B, Mistral‑7B, Gemma‑7B, DeepSeek‑7B, and Qwen2.5‑7B. All models share the same LoRA configuration and hyper‑parameters, ensuring a fair comparison.

Baseline systems include a vanilla LoRA fine‑tuning (single supervised loss) and prior A3CG baselines. Ablation studies incrementally add contrastive learning, ordinal loss, gating, and MetaGradNorm, allowing the authors to isolate the contribution of each component. The primary metric is F1 score, reported separately for seen (S) and unseen (US) categories and averaged over three folds.

Results show that the full framework (named COGLM) consistently outperforms the vanilla LoRA baseline across all models. On T5, COGLM achieves an overall F1 of 0.724, with a 4–5 point gain on the unseen categories, demonstrating markedly improved cross‑category generalization. For the 7–8 B decoder‑only models, COGLM yields 3–6 % absolute F1 improvements over LoRA‑only, and in several cases matches or exceeds the unseen‑category performance of much larger proprietary models such as GPT‑4o, Claude 3.5 Sonnet, and LLaMA‑3‑70B, despite having an order of magnitude fewer parameters.

The authors also discuss a key trade‑off: stronger ordinal constraints (larger margin m₀) increase separation of actionability levels on the training data but can lead to over‑rigid embeddings that hurt performance on unseen categories. Similarly, the gating temperatures and MetaGradNorm hyper‑parameters are sensitive; improper settings can cause instability or collapse of one loss component.

In summary, the paper makes four major contributions: (1) a structured PEFT framework that embeds contrastive and ordinal objectives directly into LoRA adapters; (2) a gated feature modulation mechanism that dynamically balances multiple objectives at the sample level; (3) MetaGradNorm, a meta‑learning based loss‑balancing strategy that stabilizes multi‑objective training; and (4) a comprehensive empirical evaluation across multiple LLMs and ESG categories, revealing that representation structure can outweigh sheer model scale for greenwashing detection. The work opens avenues for further research on automated margin learning, broader ESG domains, and interactive human‑in‑the‑loop refinement.