"I'm eating a sandwich in Glasgow"

Abstract

Social media such as Twitter generate large quantities of data about what a person is thinking and doing in a particular location. We leverage this data to build models of locations to improve our understanding of a user's geographic context. Understanding the user's geographic context in turn allows us to present information, recommend businesses and services, and place advertisements that are relevant at a hyper-local level.In this paper we create language models of locations using coordinates extracted from geotagged Twitter data. We model locations at varying levels of granularity, from zip code to the country level. We measure the accuracy of these models by the degree to which we can predict the location of an individual tweet, and further by the accuracy with which we can predict the location of a user. We find that we can meet the performance of the industry standard tool for predicting both the tweet and the user, at the country, state and city levels, and far exceed its performance at the hyper-local level, achieving a three-to ten-fold increase in accuracy at the zip code level.

On Fine-Grained Geolocalisation of Tweets

“…ey then create a vector representation of that area from the generated document using a bag-of-word approach. To geolocate a given tweet, they then nd the most similar area to that tweet based on its content-similarity, using the generated vectors [9]. Paraskevopoulos and Palpanas [13], in addition to above, have also considered time-evolution characteristics in their matching algorithm.…”

“…First, we create a grid to divide the geographical area into squares of size 1km and associate each geo-tagged tweet to an area based on its location. As discussed in Section 2, the grid approach has been widely used in the literature to represent geographical areas at di erent levels of granularity [9,13]. Second, we obtained the Top-N contentbased similar geo-tagged tweets to a non-geo-tagged using di erent retrieval models (see Section 4.1).…”

“…tweet text or user pro le -and mapping them to di erent geo-spatial datasets such as DBpedia Spotlight or Geonames. More recently, other works tackled this problem by dividing the geographical space into areas of a given size and then modelled the language for each area [9,13,17,20]. en, a ranking approach is used to retrieve the most likely area based on the probability that a non-geo-tagged tweet was issued in that area.…”

“…al. [9] by dividing the geographical space into ne-grained squares of size 1km. Also, the authors reduced the granularity of time by considering time slots of 4 hours, and computing the number of tweets by time for each candidate location compared with the global activity of the city.…”

Contextual Search: A Computational Framework

“…Finally, a combination of social variables and geographical variables is described by Kinsella et al [100]. This is another example of how language models can be exploited for modeling and integrating diverse contextual variables together.…”