On Fine-Grained Geolocalisation of Tweets

Abstract

Recently, the geolocalisation of tweets has become an important feature for a wide range of tasks in Information Retrieval and other domains, such as real-time event detection, topic detection or disaster and emergency analysis. However, the number of relevant geo-tagged tweets available remains insu cient to reliably perform such tasks. us, predicting the location of non-geotagged tweets is an important yet challenging task, which can increase the sample of geo-tagged data and help to a wide range of tasks. In this paper, we propose a location inference method that utilises a ranking approach combined with a majority voting of tweets weighted based on the credibility of its source (Twi er user). Using geo-tagged tweets from two cities, Chicago and New York (USA), our experimental results demonstrate that our method (statistically) signi cantly outperforms our baselines in terms of accuracy, and error distance, in both cities, with the cost of decrease in recall.

Learning to Geolocalise Tweets at a Fine-Grained Level

“…After the training process, we use our learned model to re-rank doc-tweets based on their probability of being posted in the same area as the query-tweet. Finally, inspired by previous work [8], we apply a majority voting algorithm to select the predicted locationa squared area of size 1km -within the Top-N doc-tweets.…”

“…Users have the option of generating a tweet sharing this information with their followers along with the geolocation of the venue. Second, following Gonzalez et al [8] approach, we compute a credibility score for the doc-tweet which represents the posting activity of the user that generated the tweet. A doc-tweet posted by a user with a high score is more likely to be representative of a geolocalisation.…”

“…Recently, Gonzalez Paule et al [8] proposed to rank individual geotagged tweets using an IDF weighting model and combine evidence from the Top-N tweets using a weighted majority voting algorithm. However, considering only IDF weighting to perform the ranking can reduce the quality of the Top-N tweets.…”

“…To tackle this drawback, more recently Gonzalez Paule et al [8] proposed to treat each geotagged tweet individually and rank them by their content-similarity to a given non-geotagged tweet. Then, a weighted majority voting algorithm is used to select the most common area -squares of size 1 km associated to each tweetwithin the Top-N most similar geotagged tweets.…”

“…However, only two sources of evidence were used to perform geolocalisation: contentbased similarity and user information. In contrast to Gonzalez Paule et al [8], we aim to increase the accuracy of geolocalisation by adopting a learning to rank approach that learns from multiple tweet features to obtain a higher quality Top-N ranking of most similar geotagged tweets that are then fed into the majority voting algorithm.…”