← All work
Analytics & ML · 2025
ESG sentiment classifier
2 of 5real-world sentences right, after 86.84% on the test set
Role
Group · I built the SVM
Year
2025
Type
Analytics & ML
Tools
NLP, scikit-learn
The answer
The model scores well on the test set but partly learned the dataset's sentence templates, not sentiment. The next step is training on real sustainability reports and news.
86.84%test accuracy, 228 sentences
2 of 5real-world sentences right
16test sentences also in training
01The problem
Can a lightweight text classifier flag whether a statement about a company's ESG practice is positive or negative?
Group project, Semester 7, 2025 · I built the SVM, one of the models the group compared; a teammate prepared the dataset
02What I did
- Built and evaluated the SVM on 910 labelled sentences (75/25 stratified split).
- Read its feature weights to see what it had learned.
- Tested it on real-world sentences.
03Results
Confusion matrix, 228 test sentences. Rows: actual label. Columns: predicted.
Source: notebook classification report (recall 0.882 x 110); notebook classification report; notebook classification report (recall 0.856 x 118)
| Sentence | Actual | Model | Right? |
|---|---|---|---|
| The company announced a new plan to achieve net-zero emissions by 2040. | Positive | Positive | yes |
| Regulators are investigating the firm over alleged human rights violations in its supply chain. | Negative | Positive | no |
| The foundation's investment in renewable energy has yielded strong community support. | Positive | Negative | no |
| Stakeholders expressed disappointment with the board's lack of diversity. | Negative | Negative | yes |
| The company's carbon footprint remains a significant concern for investors. | Negative | Positive | no |
Source: notebook cell 11
04Files
Next project
12 screens →