Addressing Subjectivity and Refining Spatial Patterns in Natural Language Processing of Mississippi River Basin Scientific Literature

Natural language processing (NLP) applications for extracting spatial information from text data require robust, reproducible, and trustworthy modeling approaches. Our study enlisted a latent Dirichlet allocation algorithm to extract latent spatial patterns in published scientific literature about the Mississippi River Basin. We focused on methods to address subjectivity and refine three “geospatial topic modeling” steps: data preparation, model calibration, and output interpretation. First, a nomenclature that categorizes text data by spatial extent mitigates risk of unexpected geographies and concepts obscuring latent topic interpretation. Second, sensitivity analysis allows for individual effects of geospatial topic model parameters to be identified, thus informing the calibration and removal of parameters necessary to refine latent topic meaningfulness. Finally, “junk topics” constructed by topic models convey language that distinguishes latent topics that possess geographical importance, thus contributing to interpretation of meaningful spatial patterns. These approaches allow us to address and mitigate subjectivity in the application of NLP for identifying spatial patterns in text data, through a combination of high-level statistics, sensitivity analysis, and human interpretation. Such techniques are applicable beyond geospatial analysis of the Mississippi River Basin’s environmental challenges, particularly as the demand for high-quality text data in NLP-based geographical studies grows in the coming years.