Automatic Identification of Expressions of Locations in Tweet Messages using Conditional Random Fields

Automatic Identification of Expressions of Locations in Tweet Messages using Conditional Random Fields Fei Liu, Afshin Rahimi, Bahar Salehi, Miji Choi, Ping Tan, Long Duong Department of Computing and Information Systems The University of Melbourne {fliu3,bsalehi,jooc1,pingt,lduong}@student.unimelb.edu.au afshinrahimi@gmail.com Abstract In this paper, we propose an automatic identification model, capable of extract- ing expressions of locations (EoLs) within Twitter messages. Moreover, we partici- pated in the competition of ALTA Shared Task 2014 and our best-performing system is ranked among the top 3 systems (2nd in the public leaderboard). In our model, we explored the validity of the use of a wide variety of lexical, structural and geospa- tial features as well as a machine learning model Conditional Random Fields (CRF). Further, we investigated the effectiveness of stacking and self-training. 1 Introduction With the rise of social media, people have devel- oped a fondness for posting not only their thoughts and opinions, but also content regarding their whereabouts (Liu et al., 2014). While the spa- tial information carried by tweets is crucial to a wide variety of location-based applications rang- ing from real-time disaster detection (Sakaki et al., 2010; N´u˜nez-Red´o et al., 2011; Yin et al., 2012) to targeted advertising (Tuten, 2008; Evans, 2012), only 26% of Twitter users specify their locations as granular as a city name (e.g. Melbourne, Aus- tralia) in their profile according to Cheng et al. (2010). Further, as little as 0.42% of all the tweets investigated by Cheng et al. (2010) are associated with the per-tweet geo-tagging feature (i.e. a lat- itude and longitude). To add to the complexity, on such highly-interactive yet informal social me- dia platforms, people make heavy use of informal language, such as acronyms (e.g. NYC) and word shortenings (e.g. St.) due to the 140-character limit (Agarwal et al., 2011; Eisenstein, 2013; Han et al., 2013), making the identification task even more difficult. Despite the difficulties, identify- ing geospatial information in social media text has drawn much attention (Lingad et al., 2013). Our focus in this paper is the automatic identi- fication of EoLs in the text of tweets consisting of any specific reference to a geospatial location. A location, as defined by Lingad et al. (2013), con- sists of both geographic location(s), such as coun- try, city, river, or suburb, and point(s)-of-interest (POI (s)) which refer to hotels, shopping centres, and restaurants. The task is closely related to Named Entity Recognition (NER). In this regard, Liu et al. (2011) and Ritter et al. (2011) report F-score of 77-78% at identifying spatial named entities in tweets. Matching place references in a gazetteer (Hill, 2000) is another widely-used approach. Pa- radesi (2011) investigated the approach of com- bining NER and external gazetteers. Further, Gel- ernter and Balaji (2013) built a geoparser incorpo- rating the results of four parsers. In our attempt to build an automatic EoL identi- fication system, we employed a conditional ran- dom field (CRF) (Lafferty et al., 2001), which can be found and has proved to be successful in various Natural Language …

Content truncated to 500 words