|
|
--- |
|
|
language: en |
|
|
license: apache-2.0 |
|
|
tags: |
|
|
- finance |
|
|
- esg |
|
|
- sentiment-analysis |
|
|
- bert |
|
|
metrics: |
|
|
- f1 |
|
|
- accuracy |
|
|
--- |
|
|
|
|
|
# finbert_esg_sentiment_classifier |
|
|
|
|
|
## Overview |
|
|
This model is a specialized BERT-based classifier fine-tuned for Environmental, Social, and Governance (ESG) sentiment analysis in financial reports. It categorizes text into specific ESG pillars or identifies neutral financial statements. |
|
|
|
|
|
|
|
|
|
|
|
## Model Architecture |
|
|
The model utilizes a **BERT-Base-Uncased** backbone with a sequence classification head. |
|
|
- **Encoder**: 12-layer Transformer. |
|
|
- **Hidden Dimensions**: 768. |
|
|
- **Head**: Linear layer followed by Softmax for 4-class categorization. |
|
|
- **Optimization**: Trained using the Cross-Entropy loss function: |
|
|
$$\mathcal{L} = -\sum_{c=1}^{M} y_{o,c} \ln(p_{o,c})$$ |
|
|
|
|
|
## Intended Use |
|
|
- **Investment Research**: Automating the extraction of ESG signals from 10-K filings and earnings transcripts. |
|
|
- **Compliance**: Monitoring corporate communications for ESG-related disclosures. |
|
|
- **Sustainable Finance**: Providing data for ESG scoring algorithms. |
|
|
|
|
|
## Limitations |
|
|
- **Context Window**: Restricted to 512 tokens. Long documents must be processed in chunks. |
|
|
- **Language**: Optimized for English financial terminology; performance on other languages or casual text is not guaranteed. |
|
|
- **Factuality**: Classification is based on linguistic patterns, not external fact-checking of the corporate claims. |
