Place-based GIS are at the forefront of GIScience research and characterized by textual descriptions, human conceptualizations as well as the spatial-semantic relationships among places. The concepts of places are difficult to handle in geographic information science and systems because of their intrinsic vagueness. They arise from the complex interaction of individuals, society, and the environment. The exact delineation of vague regions is challenging as their borders are vague and the membership within a region varies non-monotonically and as a function of context. Consequently, vague regions are difficult to handle computationally, e.g., in spatial analysis, cartography, geographic information retrieval, and GIS workflows in general. The emergence of big data brings new opportunities for us to understand the place semantics from large-scale volunteered geographic information and data streams, such as geotags, texts, activity streams, and GPS trajectories. The term "social sensing" describes such individual-level big geospatial data and the associated analysis methods. In this dissertation, I present a generalizable, data-driven framework that complements classical top-down approaches by extracting the representations of vague cognitive regions and function regions from bottom-up approaches using spatial statistics and machine learning techniques with various social sensing sources. I demonstrate how to derive crisp boundaries for cognitive and functional regions from points of interest data, and show how natural language processing techniques can enrich our understanding of places and form a foundation for the semantic characterization of place types and the generalization of regions. This work makes contributions to the development of computational methodologies for extracting vague cognitive regions and functional regions using data-driven approaches as well as the novel semantic generalization processing technique.

Place name disambiguation, i.e., toponym disambiguation or toponym resolution, is the task of correctly identifying a place from a set of places sharing a common name. It contributes to a variety of tasks such as knowledge extraction, query answering, geographic information retrieval, and automatic tagging. Disambiguation quality relies on the ability to correctly identify and interpret contextual clues, complicating the task for short texts. Here I propose a novel approach to the disambiguation of place names from short texts that integrates three models: entity co-occurrence, topic modeling, and word embedding. The first model uses Linked Data to identify related entities to improve disambiguation quality. The second model uses topic modeling to differentiate places based on the terms used to describe them. The third model uses word embeddings to uncover the semantic relatedness between places and contexts. I evaluate this approach using a corpus of short texts collected through web scraping, determine the suitable weights for the models, and demonstrate that the combined model, i.e., Things and Strings Model, outperforms benchmark systems such as DBpedia Spotlight, TextRazor, and Open Calais by up to 85% in F-score and 46% in Precision at 1. A web service is built to demonstrate the proposed method and it can be a building block for those applications that need place name recognition and disambiguation.