The French Social Media Bank: a Treebank of Noisy User Generated Content

Proceedings of COLING 2012: Technical Papers, pages 2441–2458, COLING 2012, Mumbai, December 2012. The French Social Media Bank: a Treebank of Noisy User Generated Content Djamé Seddah1,2 Benoit Sagot1 Marie Candito1 Virginie Mouilleron1 Vanessa Combet1 (1) Alpage, Inria Paris-Rocquencourt & Université Paris 7 175 rue du Chevaleret, 75013 Paris, France (2) Université Paris Sorbonne 7 rue Victor Cousin, 75006, Paris, France firstname.lastname@inria.fr ABSTRACT In recent years, statistical parsers have reached high performance levels on well-edited texts. Domain adaptation techniques have improved parsing results on text genres differing from the journalistic data most parsers are trained on. However, such corpora usually comply with standard linguistic, spelling and typographic conventions. In the meantime, the emergence of Web 2.0 communication media has caused the apparition of new types of online textual data. Although valuable, e.g., in terms of data mining and sentiment analysis, such user-generated content rarely complies with standard conventions: they are noisy. This prevents most NLP tools, especially treebank based parsers, from performing well on such data. For this reason, we have developed the French Social Media Bank, the first user-generated content treebank for French, a morphologically rich language (MRL). The first release of this resource contains 1,700 sentences from various Web 2.0 sources, including data specifically chosen for their high noisiness. We describe here how we created this treebank and expose the methodology we used for fully annotating it. We also provide baseline POS tagging and statistical constituency parsing results, which are lower by far than usual results on edited texts. This highlights the high difficulty of automatically processing such noisy data in a MRL. KEYWORDS: Treebanking, User Generated Content, Parsing, Social Media. 2441 1 Introduction Complaining about the lack of robustness of statistical parsers whenever they are applied on out-of-domain text has almost became an overused cliché over the last few years. It remains true that such parsers only perform well on texts that are comparable to their training corpus, especially in terms of genre. As noted by Foster (2010) and Foster et al. (2011b), most studies on out-of-domain statistical parsing have been focusing mainly on slightly different newspaper texts (Gildea, 2001; McClosky et al., 2006a,b), biomedical data (Lease and Charniak, 2005; Mc- Closky and Charniak, 2008) or balanced corpora mixing different genres (Foster et al., 2007). The common point between these corpora is that they are edited texts. This means that their underlying syntax, spelling, tokenization and typography remain standard, even if they slightly depart from the newspaper genre. Therefore, standard NLP tools can be used on such corpora. Now, new forms of electronic communication have emerged in the last few years,namely so- cial media and Web 2.0 communication media, either synchronous (micro-blogging) or asyn- chronous (forums), and the need for comprehensive ways of coping with the new languages types carried by those media is becoming of crucial importance. In fact, the main consequence of the Web 2.0 revolution is that what was formerly restricted to one’s inner circle of relations has now became widely available and is …

Content truncated to 500 words