Finding the Storyteller: Automatic Spoiler Tagging using Linguistic Cues

Proceedings of the 23rd International Conference on Computational Linguistics (Coling 2010), pages 412–420, Beijing, August 2010 Finding the Storyteller: Automatic Spoiler Tagging using Linguistic Cues Sheng Guo Department of Computer Science Virginia Tech guos@cs.vt.edu Naren Ramakrishnan Department of Computer Science Virginia Tech naren@cs.vt.edu Abstract Given a movie comment, does it contain a spoiler? A spoiler is a comment that, when disclosed, would ruin a surprise or reveal an important plot detail. We study automatic methods to detect comments and reviews that contain spoilers and ap- ply them to reviews from the IMDB (Inter- net Movie Database) website. We develop topic models, based on Latent Dirichlet Allocation (LDA), but using linguistic de- pendency information in place of simple features from bag of words (BOW) repre- sentations. Experimental results demon- strate the effectiveness of our technique over four movie-comment datasets of dif- ferent scales. 1 Introduction In everyday parlance, the notion of ‘spoilers’ refers to information, such as a movie plot, whose advance revelation destroys the enjoyment of the consumer. For instance, consider the movie De- railed which features Clive Owen and Jennifer Aniston. In the script, Owen is married and meets Aniston on a train during his daily commute to work. The two of them begin an affair. The adul- tery is noticed by some inscrupulous people who proceed to blackmail Owen and Aniston. To ex- perience a spoiler, consider this comment from imdb.com: I can understand why Aniston wanted to do this role, since she gets to play majorly against type (as the supposedly ‘nice’ girl who’s really - oh no! - part of the scam), but I’m at a loss to fig- ure out what Clive Owen is doing in this sub-par, unoriginal, ugly and overly violent excuse for a thriller. i.e., we learn that Aniston’s character is actually a not-so-nice person who woos married men for later blackmail, and thus a very suspenseful piece of information is revealed. Automatic ways to de- tect spoilers are crucial in large sites that host re- views and opinions. Arguably, what constitutes a spoiler is inherently a subjective assessment and, for movies/books with intricate storylines, some comments are likely to contain more spoilers than others. We therefore cast the spoiler detection problem as a ranking problem so that comments that are more likely to be spoilers are to be ranked higher than others. In particular, we rank user comments w.r.t. (i.e., given) the movie’s synopsis which, according to imdb, is ‘[a detailed description of the movie, including spoilers, so that users who haven’t seen a movie can read anything about the title]’. Our contributions are three fold. (i) We for- mulate the novel task of spoiler detection in re- views and cast it as ranking user comments against a synopsis. We demonstrate how simple bag- of-words (BOW) representations need to be aug- mented with linguistic cues in order to satisfac- torily detect spoilers. (ii) We showcase the abil- ity of dependency parses to extract discrimina- tory linguistic cues that can distinguish spoil- …

Content truncated to 500 words