Part of Speech Tagging for French Social Media Data

Proceedings of COLING 2014, the 25th International Conference on Computational Linguistics: Technical Papers, pages 1764–1772, Dublin, Ireland, August 23-29 2014. Part of Speech Tagging for French Social Media Data Farhad Nooralahzadeh ´Ecole Polytechnique de Montr´eal Montr´eal, PQ, Canada nooralahzadeh @gmail.com Caroline Brun Xerox Research Centre Europe Meylan, France caroline.brun @xrce.xerox.com Claude Roux Xerox Research Centre Europe Meylan, France claude.roux @xrce.xerox.com Abstract In the context of Social Media Analytics, Natural Language Processing tools face new chal- lenges on on-line conversational text, such as microblogs, chat, or text messages, because of the specificity of the language used in these channels. This work addresses the problem of Part- Of-Speech tagging (initially for French but also for English) on noisy language usage from the popular social media services like Twitter, Facebook and forums. We employ a linear-chain con- ditional random fields (CRFs) model, enriched with several morphological, orthographic, lexical and large-scale word clustering features. Our experiments used different feature configurations to train the model. We achieved a higher tagging performance with these features, compared to baseline results on French social media bank. Moreover, experiments on English social media content show that our model improves over previous works on these data. 1 Introduction There are many challenges inherent to applying standard natural language analysis techniques to social media. On-line conversational texts, such as tweets are quite challenging for text mining tools, and in particular for opinion mining, as they contain very little contextual information and assume too much implicit knowledge. They expose much more language variation and tend to be less grammatical than regular texts such as news articles or books. Furthermore, they contain unusual capitalization, and make frequent use of emoticons, abbreviations and hash-tags, which can form an important part of their in- ner meaning (Maynard et al., 2012). Conventional natural language processing tools for regular texts have achieved reasonably high accuracy thanks to machine learning techniques on large annotated data set. However, ”off the shelf” language processing systems fail to work on social media data and their performance on this domain degrade very fast. For example, in English Part-Of-Speech tagging, the accuracy of the Stanford tagger (Toutanova et al., 2003) falls from 97% on Wall Street Journal text to 85% accuracy on Twitter (Gimpel et al., 2011), similarly the MElt POS tagger (Denis and Sagot, 2012) drops from 97.7% on the French Treebank (called the FTB-UC by (Candito and Crabb´e, 2009)) to 85.2% on on-line conversational texts (Seddah et al., 2012). In Named Entity Recognition, the CoNLL-trained Stanford recognizer achieves 44% F-measure (Ritter et al., 2011), down from 86% on the CoNLL test set (Finkel et al., 2005); regarding parsing, see for example (Foster et al., 2011; Seddah et al., 2012), poor performances have been reported for different state-of-the-art parsers applied to English and French social media content. The main objective of this work is to implement a dedicated Part-Of-Speech (POS) tagger for French social media content such as Twitter, Facebook, blogs, forums and customer reviews. We used the first user-generated content resource …

Content truncated to 500 words