As a high-quality source in social media big data, the geographic location has been widely adopted in the fields of disease control, population mobility analysis and ad delivery positioning with the rapid development of online social media and the prevalence of localizable mobile devices. However, the location data are quite sparse because often the locations cannot be accurately specified by the users. To overcome this data sparsity problem, this paper proposes UGC-LI, a user generate content driven location inference method to infer the location where users and social texts are created. The method can provide supporting data for location-based personalized information services. A probability model is constructed by comprehensive considering the distribution of location words and social graph of users via local words extracted from user generated texts to locate the users in multi-granularity. Further, a parameterized linguistic model based on location is presented to calculate the city where the tweet is published. The results of experiment on real-word dataset demonstrate that this new method outperforms existing algorithms. In the experiment, 64.2% of users are identified within 15km displacement distance, 81.3% of the living cities and 32.7% of the cities where the tweets were tweeted are correctly located.