Identifying Humor in Reviews using Background Text Sources

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 492–501 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Identifying Humor in Reviews using Background Text Sources Alex Morales and ChengXiang Zhai Department of Computer Science University of Illinois, Urbana-Champaign amorale4@illinois.edu czhai@illinois.edu Abstract We study the problem of automatically identifying humorous text from a new kind of text data, i.e., online reviews. We propose a generative language model, based on the theory of incongruity, to model humorous text, which allows us to leverage background text sources, such as Wikipedia entry descriptions, and enables construction of multiple features for iden- tifying humorous reviews. Evaluation of these features using supervised learning for classifying reviews into humorous and non-humorous reviews shows that the fea- tures constructed based on the proposed generative model are much more effective than the major features proposed in the ex- isting literature, allowing us to achieve al- most 86% accuracy. These humorous re- view predictions can also supply good in- dicators for identifying helpful reviews. 1 Introduction The growth of online feedback systems, such as online reviews in which users can write about their preferences and opinions, has allowed for creativ- ity in the written communication of user ideas. As such, these feedback systems have become ubiq- uitous, and it’s not difficult to imagine a future with smart systems reacting to user’s behaviour in a human-like manner (Nijholt, 2014). An es- sential component for personal communication is the expression of humor. Although many peo- ple have studied the theory of humor, it still re- mains loosely defined (Ritchie, 2009), this leads to difficulties in modelling humor. While the task for identifying humor in text has been previously studied, most approaches have focused on shorter text such as Twitter data (Mihalcea and Strappar- ava, 2006; Reyes et al., 2012, 2010) (see Section 6 for a more complete review of related work). In this paper, we study the problem of automatically identifying humorous text from a new kind of text data, i.e., online reviews. In order to quantitatively test whether the review is humorous, we devise a novel approach, using the theory of incongruity, to model the reviewer’s humorous intent when writ- ing the review. The theory of incongruity states that we laugh because there is something incon- gruous (Attardo, 1994), in other words, there is a change from our expectation. Specifically, we propose a general generative language model to model the generation of humor- ous text. The proposed model is a mixture model with multinomial distributions as component mod- els (i.e., models of topics), similar to Probabilis- tic Latent Semantic Analysis (Hofmann, 1999). However, the main difference is that the compo- nent word distributions (i.e., component language models) are all assumed to be known in our model, and they are designed to model the two types of language used in a humorous text, including 1) the general background model estimated using all the reviews, and 2) the reference language models of all the topical aspects …

Content truncated to 500 words