sbic-method2

Contents

1. Labeled dataset (150 International companies, 230 CSR GRI reports, 2017-2021 period, 57k paragraphs, 360k sentences)
2. Dataset for prediction (150 German PLC companies, 1.2k CSR reports, 2010-2021 period, 645k paragraphs) (full dataset available upon request)
3. Claculated text embeddings of both datasets
4. Script to predict the labels

Multilabel Classification Steps

Prerequisites

• Python 3.8+
• Required Python libraries (install using the command below)

Input Files

1. Data Files:
   • labeled dataset: labeled.csv
   • dataset for prediction: prediction_demo.csv
2. Precomputed Embeddings:
   • labeled dataset: embeddings_labeled.pkl
   • dataset for prediction: embeddings_prediction.pkl

Running the Script

Processing Steps

1. Load Data & Pretrained Embeddings
2. Perform Cosine Similarity Search: Finds the most relevant reports (sentences) using semantic_search from sentence-transformers.
3. Apply K-Nearest Neighbor (KNN) Algorithm: Selects top similar reports (sentences) and aggregates predictions.
4. Use Sigmoid Activation for Classification: Applies a threshold to generate final classification outputs.
5. Save Results: Exports df_results_0_50k.csv containing the processed data for the first 50k of records.

Output File

Execution Time