Fine-Grained Spoiler Detection from Large-Scale Review Corpora

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2605–2610 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 2605 Fine-Grained Spoiler Detection from Large-Scale Review Corpora Mengting Wan∗, Rishabh Misra†, Ndapa Nakashole∗, Julian McAuley∗ ∗University of California, San Diego, †Amazon.com, Inc {m5wan, r1misra, nnakashole, jmcauley}@ucsd.edu Abstract This paper presents computational approaches for automatically detecting critical plot twists in reviews of media products. First, we cre- ated a large-scale book review dataset that includes fine-grained spoiler annotations at the sentence-level, as well as book and (anonymized) user information. Second, we carefully analyzed this dataset, and found that: spoiler language tends to be book-specific; spoiler distributions vary greatly across books and review authors; and spoiler sentences tend to jointly appear in the latter part of re- views. Third, inspired by these findings, we developed an end-to-end neural network ar- chitecture to detect spoiler sentences in re- view corpora. Quantitative and qualitative re- sults demonstrate that the proposed method substantially outperforms existing baselines. 1 Introduction ‘Spoilers’ on review websites can be a concern for consumers who want to fully experience the excitement that arises from the pleasurable un- certainty and suspense of media consumption (Loewenstein, 1994). Certain review websites al- low reviewers to tag whether their review (or sen- tences in their reviews) contain spoilers. However, we observe that in reality only a few users utilize this feature. Thus, requiring sentence-level spoiler annotations from users is not a successful ap- proach to comprehensive fine-grained spoiler an- notation. One possible solution is crowdsourcing: whereby consumers can report reviews that reveal critical plot details. This is complementary to the self-reporting approach, but may have scalability issues as it is relatively difficult to engage suffi- cient consumers in a timely fashion. Therefore, we seek to address the lack of completeness exhibited by self-reporting and crowdsourcing. We instead focus on developing machine learning techniques • This was a perfect, albeit bloody, end to the series. • Though there were deaths that were definitely unwarranted: <spoiler>Fred Hedwig Moody Tonks Lupin Dobby,</spoiler> there were some really heartfelt and memorable moments: <spoiler>Narcissa saving Harry, Ron coming back, Hermione and Ron, Harry and Ginny, Molly killing Bellatrix, etc.</spoiler> • I wish we could have spent more time at Hogwarts, as one of my favorite characters, the amazing Minerva McGonagall, resides there, and we couldn’t see more of her amazingness in the Battle of Hogwarts. • Harry Potter was a really, really great series that I think will be (and is) timeless. Harry Potter and the Deathly Hallows www.goodreads.com/book/show/136251 p=0.35 p=0.81 p=0.44 p=0.06 predictions from SpoilerNet review document review subject (i.e., item) review author (i.e., user) Figure 1: An example review from Goodreads, where spoiler tags and the predicted spoiler probabilities from SpoilerNet are provided. to automatically detect spoiler sentences from re- view documents. Related Work. Surprisingly, we find that spoiler analysis and detection is a relatively unexplored topic; previous work focuses on leveraging sim- ple topic models (Guo and Ramakrishnan, 2010), or …

Content truncated to 500 words