ejunprung/SASB-ESG-Category-Classifier

SASB ESG Category Classifier (Stage 2)

The SASB ESG category classifier utilizes the Universal Sentence Encoder (USE) model fine-tuned to classify ESG sentences into one of 26 Sustainability Accounting Standards Board (SASB) standards categories. Interestingly, the USE model greatly outperformed BERT in our experiments. For a full description of our training data, please refer to https://www.kaggle.com/datasets/edwardjunprung/sasb-aligned-esg-sentences.

Our classifier consists of a two-stage pipeline:

1. Stage 1 - Classify sentences as ESG or not.
2. Stage 2 - Subsequently, bucket ESG sentences into one of 26 SASB categories.

Goal

The objective is to categorize sentences within ESG documents in order to evaluate corporate ESG alignment. As an illustration, upon analyzing all sentences in Activision's annual ESG report, the SASB ESG model determined that more than 40% of sentences correspond with the Diversity & Inclusion and Human Rights SASB categories. Consequently, we can infer that Activision places a significant emphasis on these initiatives, which positions it as a potential candidate for investment funds with social impact mandates.

Model Output

The SASB ESG category classifier provides predictions for 26 different classes, which are mapped as follows:

SASB Category	Model Output
GHG Emissions	8
Air Quality	5
Energy Management	7
Water & Wastewater Management	10
Waste & Hazardous Materials Management	9
Ecological Impacts	6
Human Rights & Community Relations	23
Customer Privacy	20
Data Security	22
Access & Affordability	19
Product Quality & Safety	24
Customer Welfare	21
Selling Practices & Product Labeling	25
Labor Practices	13
Employee Health & Safety	12
Employee Engagement Diversity & Inclusion	11
Product Design & Lifecycle Management	3
Business Model Resilience	0
Supply Chain Management	4
Materials Sourcing & Efficiency	1
Physical Impacts of Climate Change	2
Business Ethics	14
Competitive Behaviour	15
Management of the Legal & Regulatory Environment	17
Critical Incident Risk Management	16
Systemic Risk Management	18

Results

Below, we present a comparison between our two-stage approach and a baseline heuristic method. The baseline method categorizes ESG sentences based solely on the presence of specific keywords. For instance, any sentence containing the phrase "human rights" would be automatically labeled under that category.

Model	Parent ESG Category	Child ESG Category
Heuristic	31%	34%
SASB ESG Model	71%	61%

Please be aware of the substantial class imbalance in our training data. For example, categories like "Competitive Behavior" are considerably less prevalent compared to others such as "GHG Emissions."

Parent Category = Environment, Social Capital, Human Capital, Business Model & Innovation, Leadership & Governance
Child Category = GHG Emissions, Air Quality, etc. Please visit https://sasb.org/standards/materiality-finder to see full list.

Misc

• Developed by: Victor Chen, Jude Zhu, Michael Liston, Edward Junprung
• Parent Model: Universal Sentence Encoder
• Blog Post: https://www.gopeaks.org/esg-mapper