Abstract

The emergence of text analytics through deep learning has unlocked a myriad of possibilities for automating administrative tasks within both corporate and governmental settings. This paper presents a novel framework designed to enhance environmental impact assessment systems. Specifically, we focus on predicting the involvement of environmental regulatory agencies in industrial projects based on project content. To tackle this challenge, we develop advanced transformers within a multilabel framework, incorporating class weights to address class imbalance. Experimental results using the Chilean environmental impact assessment system show the efficacy of the framework, achieving an excellent F1 score of 0.8729 in a 14-class multilabel scenario. By eliminating the labor-intensive manual process of inviting government agencies and allowing them to opt out of evaluating specific projects, we significantly reduced project assessment times.