Text Mining Methods for Mapping Opinions from Georeferenced Documents

With the growing availability of large volumes of textual information on the Web, text mining techniques have been gaining a growing interest. One specific text mining problem that is increasingly relevant relates to the detection of textual expressions that refer to opinions on certain topics and services. A second text mining problem, which has also been gaining a growing interest, is the identification of the geographic location that best relates to the contents of particular documents. In my MSc thesis, I empirically compared automated techniques, based on language models, for assigning documents to opinion classes and to the geospatial coordinates of latitude and longitude that best summarize their contents. Using this information, I then analyzed the possibility of building thematic maps portraying the incidence of particular classes of opinions, as extracted from documents, in different geographic areas. An extensive experimental validation has been carried out over the different components, using documents from Wikipedia and reviews from Yelp. The best performing method for geocoding textual documents combines character-based language models with a post-processing technique that uses the coordinates from the 5 most similar training documents, obtaining an average prediction error of 265 Kilometers, and a median prediction error of just 22 Kilometers. In what concerns opinion mining, analysis of opinion was done in a two-point scale schema (i.e., polarity of opinion) and a five-point scale schema (i.e., considering five degrees of opinions). The best performing methods used character-based language models, and for which the two-point scale case achieved an accuracy of 0.80. The best performing method for the five-point scale, based on a hierarchical classifier, achieved an accuracy of 0.50. A technique known as Kernel Density Estimation was used in the development of the thematic maps, and an empirical analysis has shown that the maps obtained through automatic extraction indeed correspond to an accurate representation for the geographic distribution of opinions.