End-to-end Network for Twitter Geolocation Prediction and Hashing

Proceedings of the The 8th International Joint Conference on Natural Language Processing, pages 744–753, Taipei, Taiwan, November 27 – December 1, 2017 c⃝2017 AFNLP End-to-end Network for Twitter Geolocation Prediction and Hashing Jey Han Lau1,2 Lianhua Chi1 Khoi-Nguyen Tran1 Trevor Cohn2 1 IBM Research 2 School of Computing and Information Systems, The University of Melbourne jeyhan.lau@gmail.com, lianhuac@au1.ibm.com, khndtran@au1.ibm.com, t.cohn@unimelb.edu.au Abstract We propose an end-to-end neural network to predict the geolocation of a tweet. The network takes as input a number of raw Twitter metadata such as the tweet mes- sage and associated user account informa- tion. Our model is language independent, and despite minimal feature engineering, it is interpretable and capable of learning location indicative words and timing pat- terns. Compared to state-of-the-art sys- tems, our model outperforms them by 2%- 6%. Additionally, we propose extensions to the model to compress representation learnt by the network into binary codes. Experiments show that it produces com- pact codes compared to benchmark hash- ing algorithms. An implementation of the model is released publicly.1 1 Introduction A number of applications benefit from geographi- cal information in social data, from personalised advertising to event detection to public health studies. Sloan et al. (2013) estimate that less than 1% of tweets are geotagged with their locations, motivating the development of geolocation predic- tion systems. Han et al. (2012) introduced the task of pre- dicting the location based only the tweet mes- sage. A key difference to previous work is that the prediction is made at message or tweet level, while predecessor methods tend to focus on user- level prediction (Backstrom et al., 2010; Cheng et al., 2010). Since then, various methods have been proposed for the task (Han et al., 2014; Chi et al., 2016; Jayasinghe et al., 2016; Miura et al., 1https://github.com/jhlau/ twitter-deepgeo 2016), although most systems are engineered for a particular platform and language (e.g. website- specific parsers and language-specific tokenisers and gazetteers). Another strand of research lever- ages the social network structure to infer location; Jurgens et al. (2015) provided a standardised com- parison of these systems. Our focus in this pa- per is on using only the tweets, although Rahimi et al. (2015) showed that the best approach maybe to combine both types of information. In applications where fast retrieval of co-located tweets is necessary (e.g. disaster detection), effi- cient representation of large volume of tweets con- stitute an important issue. Traditionally, hashing techniques such as locality sensitive hashing (In- dyk and Motwani, 1998) are used to compress data into binary codes for fast retrieval (e.g. with multi- index hash tables (Norouzi et al., 2012, 2014)), but it is not immediately clear how they can interact with raw Twitter metadata — as they often require a vector as input — and incorporate supervision.2 To this end, we propose an end-to-end neu- ral network for tweet-level geolocation prediction. Our network is designed to be interpretable: we show it has the capacity to automatically learn lo- cation indicative words and activity patterns from …

Content truncated to 500 words