Abstract:
There is a growing need for organisations to align their operations and services with the sustainable development goals. In Kenya and most countries across the globe, it is mandatory requirement, for organisations especially hose trading in the stock exchange, to provide their sustainability reports annually. The categorisation and classification process, however, for the reports has become very complex and demanding especially given the growing number of reports and the diverse nature of the data and information contained in them. Traditional categorisation process has proven inadequate, inefficient and not very accurate as compared to the modern process which integrates artificial intelligence and multimodal deep learning. Due to the aforementioned, this study aimed at developing an artificial intelligent model that can read various data formats, including textual, graphical and numeric and ensure capture the varied and intricate facts with accuracy and precision.The study further sort to conduct an analysis on the integration of the multimodal data from sustainability reports and understand its impacts on categorization accuracy and provides deeper ESG insights. The findings of the study informed the development of an AI model that automated the categorisation of the sustainability reports through the employing the Transformer Models, Recurrent Neural Networks (RNN), Convolutional Neural Networks (CNN), and Fully Connected Neural Networks (FCNN), in conjunction with high-performance computing resources. Approximately, 50 sustainability reports drawn from the submissions made by the organisations that are trading in Nairobi stock exchange and the Global Reporting Initiative database, were used to test the model. The results confirmed the model's robust performance, achieving an overall accuracy of 79.0%, an F1-score of 0.76, precision of 0.82, and a recall of 0.77. This performance demonstrated a significant enhancement in efficacy and accuracy compared to traditional unimodal baselines, thereby addressing the issue of inefficiency and cost of manual analysis. The model was also enhanced to extract key metadata, check publication dates, title company, and country, and provide percentages of alignment of the organizations' processes to the Sustainable Development Goals (SDGs). The model was designed to accommodate different types of deep learning models, thus can analyse images, texts and numerical. The findings indicate the significance of multimodal learning and advanced deep learning models in enhancing ESG reporting quality and competence to make evidence-based judgments in the sustainability sector.
