A Dependency Parser for Tweets

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1001–1012, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics A Dependency Parser for Tweets Lingpeng Kong Nathan Schneider Swabha Swayamdipta Archna Bhatia Chris Dyer Noah A. Smith Language Technologies Institute Carnegie Mellon University Pittsburgh, PA 15213, USA {lingpenk,nschneid,swabha,archna,cdyer,nasmith}@cs.cmu.edu Abstract We describe a new dependency parser for English tweets, TWEEBOPARSER. The parser builds on several contributions: new syntactic annotations for a corpus of tweets (TWEEBANK), with conventions informed by the domain; adaptations to a statistical parsing algorithm; and a new approach to exploiting out-of-domain Penn Treebank data. Our experiments show that the parser achieves over 80% unlabeled attachment accuracy on our new, high-quality test set and measure the benefit of our contribu- tions. Our dataset and parser can be found at http://www.ark.cs.cmu.edu/TweetNLP. 1 Introduction In contrast to the edited, standardized language of traditional publications such as news reports, social media text closely represents language as it is used by people in their everyday lives. These informal texts, which account for ever larger proportions of written content, are of considerable interest to re- searchers, with applications such as sentiment anal- ysis (Greene and Resnik, 2009; Kouloumpis et al., 2011). However, their often nonstandard content makes them challenging for traditional NLP tools. Among the tools currently available for tweets are a POS tagger (Gimpel et al., 2011; Owoputi et al., 2013) and a named entity recognizer (Ritter et al., 2011)—but not a parser. Important steps have been taken. The English Web Treebank (Bies et al., 2012) represents an annotation effort on web text—which likely lies somewhere between newspaper text and social me- dia messages in formality and care of editing—that was sufficient to support a shared task (Petrov and McDonald, 2012). Foster et al. (2011b) annotated a small test set of tweets to evaluate parsers trained on the Penn Treebank (Marcus et al., 1993), aug- mented using semi-supervision and in-domain data. Others, such as Soni et al. (2014), have used exist- ing Penn Treebank–trained models on tweets. In this work, we argue that the Penn Treebank approach to annotation—while well-matched to edited genres like newswire—is poorly suited to more informal genres. Our starting point is that rapid, small-scale annotation efforts performed by imperfectly-trained annotators should provide enough evidence to train an effective parser. We see this starting point as a necessity, given observa- tions about the rapidly changing nature of tweets (Eisenstein, 2013), the attested difficulties of do- main adaptation for parsing (Dredze et al., 2007), and the expense of creating Penn Treebank–style annotations (Marcus et al., 1993). This paper presents TWEEBOPARSER, the first syntactic dependency parser designed explicitly for English tweets. We developed this parser follow- ing current best practices in empirical NLP: we annotate a corpus (TWEEBANK) and train the pa- rameters of a statistical parsing algorithm. Our research contributions include: • a survey of key challenges posed by syntactic analysis of tweets (by humans or machines) and decisions motivated by those challenges and by …

Content truncated to 500 words