TwitIE: An Open-Source Information Extraction Pipeline for Microblog Text

Proceedings of Recent Advances in Natural Language Processing, pages 83–90, Hissar, Bulgaria, 7-13 September 2013. TwitIE: An Open-Source Information Extraction Pipeline for Microblog Text Kalina Bontcheva, Leon Derczynski, Adam Funk, Mark A. Greenwood, Diana Maynard, Niraj Aswani University of Sheffield Initial.Surname@dcs.shef.ac.uk Abstract Twitter is the largest source of microblog text, responsible for gigabytes of human discourse every day. Processing microblog text is diffi- cult: the genre is noisy, documents have lit- tle context, and utterances are very short. As such, conventional NLP tools fail when faced with tweets and other microblog text. We present TwitIE, an open-source NLP pipeline customised to microblog text at every stage. Additionally, it includes Twitter-specific data import and metadata handling. This paper in- troduces each stage of the TwitIE pipeline, which is a modification of the GATE ANNIE open-source pipeline for news text. An eval- uation against some state-of-the-art systems is also presented. 1 Introduction Researchers have started recently to study the prob- lem of mining social media content automatically (e.g. (Rowe et al., 2013; Nagarajan and Gamon, 2011; Farzindar and Inkpen, 2012; Bontcheva and Rout, 2013)). The focus of this paper is on information ex- traction, but other active topics include opinion min- ing (Maynard et al., 2012; Pak and Paroubek, 2010), summarisation (e.g. (Chakrabarti and Punera, 2011)), and visual analytics and user and community mod- elling (Bontcheva and Rout, 2013). Social media min- ing is relevant in many application contexts, includ- ing knowledge management, competitor intelligence, customer relation management, eHealth, and eGovern- ment. Information extraction from social media content has only recently become an active research topic, following early experiments which showed this genre to be extremely challenging for state-of-the-art algo- rithms (Derczynski et al., 2013a). Simple domain adaptation techniques (e.g. (Daum´e and Marcu, 2007) are not so useful on this genre, in part due to its un- usual structure and representation of discourse, which can switch between one-to-one conversation, multi- party conversation and broadcast messages. For in- stance, named entity recognition methods typically have 85-90% accuracy on longer texts, but 30-50% on tweets (Ritter et al., 2011; Liu et al., 2012). This paper introduces the TwitIE information extraction system, which has been specifically adapted to microblog content. It is based on the most recent GATE (Cunningham et al., 2013) algorithms and is available as a GATE plugin available to download from https://gate.ac.uk/wiki/twitie.html, usable both via the GATE Developer user interface and via the GATE API. Comparisons against other state-of-the-art research on this topic are also made. 2 Related Work In terms of Named Entity Recognition (NER), and In- formation Extraction (IE) in general, microblogs are possibly the hardest kind of content to process. First, their shortness (maximum 140 characters for tweets) makes them hard to interpret. Consequently, ambigu- ity is a major problem since IE methods cannot easily make use of coreference information. Unlike longer news articles, there is a low amount of discourse infor- mation per microblog document, and threaded struc- ture is fragmented across multiple documents, flowing in multiple directions. …

Content truncated to 500 words