Periodic country analyses elaborated by the Famine Early Warning Systems Network of the United States Agency for International Development (FEWS NET) comprises a unique source of knowledge encompassing consistent reporting in over two dozen countries. This paper proposes a systematic assessment of this documentation to provide a comprehensive historical overview of food insecurity in FEWS NET covered countries. We apply an integrated machine learning approach, particularly through text mining techniques, to analyse reports. Results show a wide heterogeneity in topic prevalence, both at the temporal and geographical scales. Overall, the evidence shows that advances in machine learning and big data research offer great potential for international development agencies to leverage the vast information generated from reports to gain new insights, providing analytics that can support and improve decision-making.