ESG Document Evaluation Using Embeddings and Label Similarity
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for evaluating Environmental, Social, and Governance (ESG) policies in investment contexts rely heavily on manual analysis of unstructured documents, lacking flexibility and efficiency, especially when investment frameworks change.
Innovation Solution
A computer-implemented method using natural language processing (NLP) and graph models to automatically evaluate ESG documents by generating sentence and label embeddings, calculating similarity measures, and creating a network to determine ESG scores with minimal human supervision.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual analysis of unstructured documents is used to evaluate ESG policies, then analysts can incorporate information into financial models, but the process requires significant manual effort and time, especially when frameworks need to be modified
Solution Approach 1:
The patent replaces the mechanical manual analysis system with an automated NLP-based system. Machine learning models process unstructured ESG documents, extract relevant information, and generate embeddings that are automatically compared against framework labels, eliminating the need for manual reading and analysis of thousands of documents while maintaining evaluation accuracy
Solution Approach 2:
The system enables self-service evaluation by automatically processing documents and aligning them with ESG frameworks. The automated pipeline includes document ingestion, sentence embedding generation, label embedding creation, similarity calculation, and score generation without requiring human analysts to manually review each document, thus dramatically reducing time loss
2Adaptability or versatility
If manual work is performed to analyze ESG documents, then information can be incorporated into financial models, but the work needs to be repeated each time the investment framework is modified
Solution Approach 1:
The patent implements a dynamic system where the NLP pipeline can automatically adapt to framework modifications. When ESG frameworks are updated, the system can reprocess documents with new labels and embeddings without requiring complete manual reanalysis, as the automated pipeline dynamically adjusts to framework changes while maintaining processing efficiency
Solution Approach 2:
The automated NLP system serves multiple functions: it processes unstructured documents, generates embeddings, compares them against framework labels, and produces ESG scores. This universal system handles both initial evaluations and framework updates through the same automated pipeline, improving productivity while maintaining adaptability to different framework versions
3Loss of information
If unstructured documents are analyzed without appropriate labels, then all document information is available, but alignment with investment frameworks becomes difficult
Solution Approach 1:
The patent introduces embeddings as an intermediary representation between unstructured document text and structured framework labels. The NLP model converts document sentences into vector embeddings, which are then compared against label embeddings using similarity measures. This intermediary representation preserves all document information while enabling automated framework alignment through mathematical similarity calculations
Solution Approach 2:
The system transforms unstructured text into a different parameter space using embeddings. By converting sentences and labels into vector representations, the system changes the parameter form from raw text to numerical vectors, enabling automated similarity measurement and framework alignment while retaining the full information content of the original documents
Data Source
AI summary
A method for evaluating electronic documents based on a framework having a plurality of labels includes receiving a plurality of electronic documents, each electronic document having a company identifier. The method includes extracting a plurality of input sentences from each electronic document. The method includes generating a plurality of input sentence embeddings, each input sentence embedding corresponding to one of the input sentences. The method includes generating a plurality of label embeddings, each label embedding corresponding to one of the labels. The method includes calculating a plurality of similarity measures between each of the input sentence embeddings and each of the label embeddings. The method includes generating a plurality of label sentences for each of the labels based on the similarity measures, each of the label sentences being associated with one of the company identifiers. The method includes generating a plurality of entities for each of the label sentences.