NLP-Driven Systems in AI-Powered ESG Information Systems

This language is not only critical for investment stakeholders to evaluate and/or comply with and/or compare ESG practices but also influences how NLP models identify items that data systems track and analyze—concept identification & extraction. How things relate to one another—clustering and classification—makes its way into programs and data designs, influences AI training strategies and outputs, and ultimately dictates whether the data exploited is actually of value to the organization.

From straightforward knowledge management to sophisticated AI models, ontologies have proved great potential in capturing expertise while being particularly apposite to today’s data abundance and digital transformation. An ontology is generally intended to act as a standard—sort of common language—forming a set of controlled vocabularies and concepts. Overall, these ontologies bring opportunity to ESG teams to have control on algorithms and data they use, paving the way to create resilient ESG planning models, increase productivity, and deliver effective guidance to investment strategies. Investors can thus embrace a data-driven approach, expanding the boundaries and reshaping traditional patterns.

We use dynamic ontologies to model evolving ESG concepts and taxono¬mies29, which are aimed to facilitate the alignment and exchanging of ESG data, such as ESG material aspects, SDG indi¬cators30, multi-provider risk scores, etc., and the development of powerful ESG information systems. Classification techniques ranging from complex language models to simple sentiment analysis can be then used on top of these ontology-based models to enrich the data and the analytics process.

Sentiment analysis is one of the common NLP topics that investors can start implementing easily, especially because it has low complexity, requires manageable data and fits with a large number of ESG applications31. For example, sentiment analysis can be integrated to the risk and opportunity analytics as a way to help screen early signals. It can also support portfolio monitoring, engagement strategies and marketing initiatives.

From a technical perspective, sentiment analysis falls into the broad category of supervised learning text classification, where the model inputs a sentence and outputs a score for each sentiment class (the number of classes can vary, but it is common to use two to three simple classes: positive, neutral and negative). One popular and simple technique for developing sentiment analysis models is to use a bag-of-words representation that transforms sentences into vectors of weighted words.

To understand more about implementing an AI-powered ESG information system, download our White Paper on this subject.