Knowledge mining from scientific literature for acute aquatic toxicity: classification for hybrid predictive modelling

Abstract

This work proposes a systematic method consisting of state-of-the-art text processing approaches and human-machine interaction for the extraction of useful sentences and data in tabular, graphical, and numerical form, containing information particularly relevant for hybrid modelling. It is applied to the domain of acute aquatic toxicity of chemicals, which is particularly relevant for the safety, health, and environmental hazard assessment of chemicals. Nearly 400 papers from 2000-2021 were identified and processed with the proposed method. The results indicate that the vast amount of knowledge can be efficiently processed in orders of magnitude faster than conventional methods without loss of detail and interpretation depth. The information is in a form that can be useful in hybrid modelling with respect to model and predictor selection, prioritization, and constraints, addressing data gaps, and validating and interpreting model performance.