Classification of Poverty Condition Using Natural Language Processing

This work introduces a methodology to classify between poor and extremely poor people through Natural Language Processing. The approach serves as a baseline to understand and classify poverty through the people’s discourses using machine learning algorithms. Based on classical and modern word vector representations we propose two strategies for document level representations: (1) document-level features based on the concatenation of descriptive statistics and (2) Gaussian mixture models. Three classification methods are systematically evaluated: Support Vector Machines, Random Forest, and Extreme Gradient Boosting. The fourth best experiments yielded around 55% of accuracy, while the embeddings based on GloVe word vectors yielded a sensitivity of 79.6% which could be of great interest for the public policy makers to accurately find people who need to be prioritized in social programs.