@inproceedings{raspanti-etal-2026-greench,
title = "Greench-v1: distilling {SLM}s on Greenwashing Detection",
author = "Raspanti, Federico and
Bardelli, Alessandro Pietro Bardelli and
Scala, Simona and
Demirta{\c{s}}, {\.I}rem and
Di Bari, Marilena and
Filannino, Michele",
editor = "Grasso, Francesca and
Basile, Valerio and
Bosco, Cristina and
Ibrohim, Muhammad Okky and
Skeppstedt, Maria and
Stede, Manfred",
booktitle = "Proceedings of the 2nd Workshop on Ecology, Environment, and Natural Language Processing",
month = may,
year = "2026",
address = "Palma de Mallorca, Spain",
publisher = "European Language Resources Association",
url = "https://aclanthology.org/2026.nlp4ecology-1.8/",
doi = "10.63317/2ngqn2gr78du",
pages = "79--86",
abstract = "Validating greenwashing claims in environmental, social, and governance (ESG) reports relies heavily on costly and inconsistent manual review. To address this, this paper introduces Greench-v1, a low-latency small language model (based on Qwen3-4B) that screens ESG text at the paragraph level. The model outputs a three-way classification (Greenwashing Alert, No Greenwashing, Not Relevant) paired with a concise, paragraph-grounded rationale to assist human auditors in triage and validation. The system was trained on a custom dataset of roughly 2,000 paragraphs, adapted from the ClimateBERT corpus. This dataset mitigates class imbalance through controlled paraphrasing of rare positive instances and uses GPT-4o to generate evidence-based justifications. Four training regimes were evaluated: (i) Hard distillation: Supervised fine-tuning on teacher-generated outputs. (ii) Soft distillation: Training the student to match the temperature-scaled logits of a domain-specialized Qwen3-14B teacher. (iii) Group Relative Policy Optimization (GRPO): Reward-based updates driven by exact-match alert generation. (iv) Hybrid GRPO: GRPO initialized from the hard-distilled checkpoint. Distillation and efficient policy optimization significantly improved performance over untuned baselines. Soft distillation and GRPO achieved the strongest results, increasing the ``Greenwashing Alert'' weighted F1-score by 36.7{\%} and 49.0{\%}, respectively, resulting in a deployable tool for screening large volumes of ESG narratives."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="raspanti-etal-2026-greench">
<titleInfo>
<title>Greench-v1: distilling SLMs on Greenwashing Detection</title>
</titleInfo>
<name type="personal">
<namePart type="given">Federico</namePart>
<namePart type="family">Raspanti</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alessandro</namePart>
<namePart type="given">Pietro</namePart>
<namePart type="given">Bardelli</namePart>
<namePart type="family">Bardelli</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Simona</namePart>
<namePart type="family">Scala</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">İrem</namePart>
<namePart type="family">Demirtaş</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Marilena</namePart>
<namePart type="family">Di Bari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Michele</namePart>
<namePart type="family">Filannino</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-05</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2nd Workshop on Ecology, Environment, and Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Francesca</namePart>
<namePart type="family">Grasso</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Valerio</namePart>
<namePart type="family">Basile</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Cristina</namePart>
<namePart type="family">Bosco</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Muhammad</namePart>
<namePart type="given">Okky</namePart>
<namePart type="family">Ibrohim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Maria</namePart>
<namePart type="family">Skeppstedt</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Manfred</namePart>
<namePart type="family">Stede</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>European Language Resources Association</publisher>
<place>
<placeTerm type="text">Palma de Mallorca, Spain</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Validating greenwashing claims in environmental, social, and governance (ESG) reports relies heavily on costly and inconsistent manual review. To address this, this paper introduces Greench-v1, a low-latency small language model (based on Qwen3-4B) that screens ESG text at the paragraph level. The model outputs a three-way classification (Greenwashing Alert, No Greenwashing, Not Relevant) paired with a concise, paragraph-grounded rationale to assist human auditors in triage and validation. The system was trained on a custom dataset of roughly 2,000 paragraphs, adapted from the ClimateBERT corpus. This dataset mitigates class imbalance through controlled paraphrasing of rare positive instances and uses GPT-4o to generate evidence-based justifications. Four training regimes were evaluated: (i) Hard distillation: Supervised fine-tuning on teacher-generated outputs. (ii) Soft distillation: Training the student to match the temperature-scaled logits of a domain-specialized Qwen3-14B teacher. (iii) Group Relative Policy Optimization (GRPO): Reward-based updates driven by exact-match alert generation. (iv) Hybrid GRPO: GRPO initialized from the hard-distilled checkpoint. Distillation and efficient policy optimization significantly improved performance over untuned baselines. Soft distillation and GRPO achieved the strongest results, increasing the “Greenwashing Alert” weighted F1-score by 36.7% and 49.0%, respectively, resulting in a deployable tool for screening large volumes of ESG narratives.</abstract>
<identifier type="citekey">raspanti-etal-2026-greench</identifier>
<identifier type="doi">10.63317/2ngqn2gr78du</identifier>
<location>
<url>https://aclanthology.org/2026.nlp4ecology-1.8/</url>
</location>
<part>
<date>2026-05</date>
<extent unit="page">
<start>79</start>
<end>86</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Greench-v1: distilling SLMs on Greenwashing Detection
%A Raspanti, Federico
%A Bardelli, Alessandro Pietro Bardelli
%A Scala, Simona
%A Demirtaş, İrem
%A Di Bari, Marilena
%A Filannino, Michele
%Y Grasso, Francesca
%Y Basile, Valerio
%Y Bosco, Cristina
%Y Ibrohim, Muhammad Okky
%Y Skeppstedt, Maria
%Y Stede, Manfred
%S Proceedings of the 2nd Workshop on Ecology, Environment, and Natural Language Processing
%D 2026
%8 May
%I European Language Resources Association
%C Palma de Mallorca, Spain
%F raspanti-etal-2026-greench
%X Validating greenwashing claims in environmental, social, and governance (ESG) reports relies heavily on costly and inconsistent manual review. To address this, this paper introduces Greench-v1, a low-latency small language model (based on Qwen3-4B) that screens ESG text at the paragraph level. The model outputs a three-way classification (Greenwashing Alert, No Greenwashing, Not Relevant) paired with a concise, paragraph-grounded rationale to assist human auditors in triage and validation. The system was trained on a custom dataset of roughly 2,000 paragraphs, adapted from the ClimateBERT corpus. This dataset mitigates class imbalance through controlled paraphrasing of rare positive instances and uses GPT-4o to generate evidence-based justifications. Four training regimes were evaluated: (i) Hard distillation: Supervised fine-tuning on teacher-generated outputs. (ii) Soft distillation: Training the student to match the temperature-scaled logits of a domain-specialized Qwen3-14B teacher. (iii) Group Relative Policy Optimization (GRPO): Reward-based updates driven by exact-match alert generation. (iv) Hybrid GRPO: GRPO initialized from the hard-distilled checkpoint. Distillation and efficient policy optimization significantly improved performance over untuned baselines. Soft distillation and GRPO achieved the strongest results, increasing the “Greenwashing Alert” weighted F1-score by 36.7% and 49.0%, respectively, resulting in a deployable tool for screening large volumes of ESG narratives.
%R 10.63317/2ngqn2gr78du
%U https://aclanthology.org/2026.nlp4ecology-1.8/
%U https://doi.org/10.63317/2ngqn2gr78du
%P 79-86
Markdown (Informal)
[Greench-v1: distilling SLMs on Greenwashing Detection](https://aclanthology.org/2026.nlp4ecology-1.8/) (Raspanti et al., NLP4Ecology 2026)
ACL
- Federico Raspanti, Alessandro Pietro Bardelli Bardelli, Simona Scala, İrem Demirtaş, Marilena Di Bari, and Michele Filannino. 2026. Greench-v1: distilling SLMs on Greenwashing Detection. In Proceedings of the 2nd Workshop on Ecology, Environment, and Natural Language Processing, pages 79–86, Palma de Mallorca, Spain. European Language Resources Association.