Identifying Twitter Location Mentions

Identifying Twitter Location Mentions Bo Han, Antonio Jimeno Yepes, Andrew MacKinlay, Qiang Chen IBM Research Melbourne, VIC, Australia bohan.ibm@au1.ibm.com, antonio.jimeno@au1.ibm.com, admackin@au1.ibm.com, qiangchen@au1.ibm.com Abstract This paper describes our system in the ALTA shared task 2014. The task is to identify location mentions in Twitter mes- sages, such as place names and point-of- interests (POIs). We formulated the task as a sequential labelling problem, and ex- plored various features on top of a condi- tional random field (CRF) classifier. The system achieved 0.726 mean-F measure on the held-out evaluation data. We dis- cuss our results and suggest ideas for fu- ture work on location mention recognition in social media. 1 Introduction The ALTA shared task 2014 aims to identify loca- tion mentions in Twitter data. The input is plain text messages, and the expected output is loca- tion entities such as country names, city names and POIs for each message. For instance, Auck- land and #eqnz are identified as location mentions in @USER are you considering an Auckland visit after #eqnz today?.1 This shared task is very simi- lar to a well-established NLP task — named entity recognition (NER) but with a focus on location en- tities in social media. Each token in a text message is categorised as either a location mention or not. The nearby tokens (i.e., context) may influence a token’s labelling, hence we incorporate context in- formation in our system. Following the literature on NER (Lingad et al., 2013), we formulate it as a sequential labelling task and use a conditional random field (CRF) as the classifier. The main contributions of the paper are: (1) A sequential labeller for identifying location men- tions in social media; (2) Feature analysis and comparison in NER between social media and 1#eqnz is a short form for earthquake in New Zealand. other genres. (3) Discussion on errors and exten- sions to current sequential labeller. 2 Challenges Although CRF models for NER are widely used and are reported to achieve state-of-the-art results in literature (Finkel et al., 2005; Liu et al., 2011; Ritter et al., 2011), NER in social media still raises several non-trivial challenges. First, Twitter text is noisy, with more non- standard words than polished text (Baldwin et al., 2013) including typos (e.g., challanges “challenges”), abbreviations (e.g., ppl “people”) and phonetic substitutions (e.g., 4eva “forever”). These non-standard words often cause generali- sation issues (Han and Baldwin, 2011). For in- stance, lexical variants (e.g., Melb, Mel, melbn) will not be recognised in the test data when only standard forms (e.g., “Melbourne”) are observed in the training data. In addition to non-standard words, informal writing style further reduces NER accuracy. One example is that conventional features relying on capitalisation are less reliable. For instance, LOL is capitalised but it is not a location entity, while brisbane may be a valid location mention even though it is in lowercase. Similarly, Twitter specific entities sometimes are sentence constituents, e.g., #Melbourne in #Melbourne is my fav city. However, they may be a topic tag that does …

Content truncated to 500 words