Knowledge rich approach of processing documents has been viewed as a method to improve over simple bag-of- word representation. Extracting location information from documents and link them to some ontology such as world gazetteer through a disambiguation process becomes an interesting and important topic. Lacking of training data is a problem in disambiguation method. In this paper we described a method to automatically extract training data from large collection of documents based on local context disambiguation, and then sense profiles are generated automatically for disambiguation use. Another topic of this paper is to describe a linear combination method to combine different types of evidences of disambiguation. We explored three different evidences including location sense context in training documents, local neighbor context, and the popularity of individual location sense. Our results show that combining the three evidences generates reasonable results.