fsssg/Biodiversity_Sentiment_Classifier

Model overview

This model is a multi-class text classifier designed to analyse the sentiment of biodiversity-related content in corporate sustainability reports at the paragraph level.

The model distinguishes between three sentiment classes:

Positive (label = 0): content highlighting business opportunities related to biodiversity, positive biodiversity impacts, mitigation of negative impacts, or using positive framing and evaluative language

Neutral (label = 1): factual or descriptive statements about biodiversity without positive or negative framing, including objective reporting of data, statistics, or contextual information

Negative (label = 2): content describing actual or potential risks, losses, adverse impacts on the firm, the firm’s negative impacts on biodiversity, or using negative framing

The model is intended for research use in the analysis of corporate sustainability and ESG disclosures.

Training approach

The model was trained on a curated dataset of 2,000 manually annotated paragraphs extracted from sustainability reports of Fortune Global 500 companies.

The classifier is based on climatebert/distilroberta-base-climate-sentiment, a DistilRoBERTa-based ClimateBERT model pre-trained on climate-related corpora and previously fine-tuned for environmental sentiment classification. This model was further fine-tuned to capture sentiment in biodiversity-related corporate disclosures.

Key training characteristics include:

unit of analysis: paragraph

maximum sequence length: 256 tokens

task: multi-class sequence classification (three classes)

loss function: cross-entropy

optimisation: supervised fine-tuning using the Hugging Face Trainer API

training regime: 5-fold stratified cross-validation

Training was performed on CPU using fixed hyperparameters selected prior to cross-validation. The released model checkpoint corresponds to the fold achieving the highest weighted F1 score.

Evaluation

Performance is reported as averages across 5-fold stratified cross-validation on the annotated dataset:

weighted F1 score: 0.901

weighted precision: 0.902

weighted recall: 0.903

AUC–ROC (one-vs-rest): 0.968

Weighted metrics are reported to account for class imbalance across sentiment categories. Weighted F1 was used as the model selection criterion.

Recommended Pipeline First, use ESGBERT/EnvironmentalBERT-biodiversity to identify biodiversity-related paragraphs, then apply this model to identify commitments.