Opinion Mining on YouTube

Proceedings of the 52nd Annual Meeting of the Association for Computational Linguistics, pages 1252–1261, Baltimore, Maryland, USA, June 23-25 2014. c⃝2014 Association for Computational Linguistics Opinion Mining on YouTube Aliaksei Severyn1, Alessandro Moschitti3,1, Olga Uryupina1, Barbara Plank2, Katja Filippova4 1DISI - University of Trento, 2CLT - University of Copenhagen, 3Qatar Computing Research Institute, 4Google Inc. severyn@disi.unitn.it, amoschitti@qf.org.qa, uryupina@gmail.com, bplank@cst.dk, katjaf@google.com Abstract This paper defines a systematic approach to Opinion Mining (OM) on YouTube comments by (i) modeling classifiers for predicting the opinion polarity and the type of comment and (ii) proposing ro- bust shallow syntactic structures for im- proving model adaptability. We rely on the tree kernel technology to automatically ex- tract and learn features with better gener- alization power than bag-of-words. An ex- tensive empirical evaluation on our manu- ally annotated YouTube comments corpus shows a high classification accuracy and highlights the benefits of structural mod- els in a cross-domain setting. 1 Introduction Social media such as Twitter, Facebook or YouTube contain rapidly changing information generated by millions of users that can dramati- cally affect the reputation of a person or an orga- nization. This raises the importance of automatic extraction of sentiments and opinions expressed in social media. YouTube is a unique environment, just like Twitter, but probably even richer: multi-modal, with a social graph, and discussions between peo- ple sharing an interest. Hence, doing sentiment research in such an environment is highly relevant for the community. While the linguistic conven- tions used on Twitter and YouTube indeed show similarities (Baldwin et al., 2013), focusing on YouTube allows to exploit context information, possibly also multi-modal information, not avail- able in isolated tweets, thus rendering it a valuable resource for the future research. Nevertheless, there is almost no work showing effective OM on YouTube comments. To the best of our knowledge, the only exception is given by the classification system of YouTube comments proposed by Siersdorfer et al. (2010). While previous state-of-the-art models for opin- ion classification have been successfully applied to traditional corpora (Pang and Lee, 2008), YouTube comments pose additional challenges: (i) polarity words can refer to either video or prod- uct while expressing contrasting sentiments; (ii) many comments are unrelated or contain spam; and (iii) learning supervised models requires train- ing data for each different YouTube domain, e.g., tablets, automobiles, etc. For example, consider a typical comment on a YouTube review video about a Motorola Xoom tablet: this guy really puts a negative spin on this , and I ’m not sure why , this seems crazy fast , and I ’m not entirely sure why his pinch to zoom his laggy all the other xoom reviews The comment contains a product name xoom and some negative expressions, thus, a bag-of-words model would derive a negative polarity for this product. In contrast, the opinion towards the prod- uct is neutral as the negative sentiment is ex- pressed towards the video. Similarly, the follow- ing comment: iPad 2 is better. the superior apps just destroy the xoom. …

Content truncated to 500 words