A myriad of textual documents produced/consumed by UNEP need to be mapped to the SDGs (project proposals, reports, briefings, etc.). Such mapping exercises are time consuming for experts and rely on personal knowledge of the links between topics and the SDGs. While UNEP has experts in many fields, links to the SDGs that are outside our expertise can be overlooked. Our SDG Meter web platform is proposed as a tool to help analyze text via a neural network-based technique for natural language processing (NLP) pre-training that assesses the relationship to each of the 17 SDGs.
This technique is based on the BERT (Bidirectional Encoder Representations from Transformers) model developped by Google researchers (see paper here) with the use of the multilabel text classification feature. Our algorithm has been trained with about 3000 texts and labels extracted from the categories "News", "Guest articles" and "Policy Briefs" of the IISD-SDG website. Our method has an accuracy of 98% on 500, which means that for 500 test texts our method correctly classifies 490 texts. To have an idea of all the capacities of BERT and also its functioning explained in a very simple way we recommend you this short video.
