Duluth at SemEval-2017 Task 6: Language Models in Humor Detection

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 385–389, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics Duluth at SemEval-2017 Task 6: Language Models in Humor Detection Xinru Yan & Ted Pedersen Department of Computer Science University of Minnesota Duluth Duluth, MN, 55812 USA {yanxx418,tpederse}@d.umn.edu Abstract This paper describes the Duluth system that participated in SemEval-2017 Task 6 #HashtagWars: Learning a Sense of Hu- mor. The system participated in Subtasks A and B using N-gram language models, ranking highly in the task evaluation. This paper discusses the results of our system in the development and evaluation stages and from two post-evaluation runs. 1 Introduction Humor is an expression of human uniqueness and intelligence and has drawn attention in di- verse areas such as linguistics, psychology, phi- losophy and computer science. Computational humor draws from all of these fields and is a relatively new area of study. There is some history of systems that are able to gener- ate humor (e.g., (Stock and Strapparava, 2003), ( ¨Ozbal and Strapparava, 2012)). However, hu- mor detection remains a less explored and chal- lenging problem (e.g., (Mihalcea and Strapparava, 2006), (Zhang and Liu, 2014), (Shahaf et al., 2015), (Miller and Gurevych, 2015)). SemEval-2017 Task 6 (Potash et al., 2017) also focuses on humor detection by asking participants to develop systems that learn a sense of humor from the Comedy Central TV show, @midnight with Chris Hardwick. Our system ranks tweets ac- cording to how funny they are by training N-gram language models on two different corpora. One consisting of funny tweets provided by the task organizers, and the other on a freely available re- search corpus of news data. The funny tweet data is made up of tweets that are intended to be hu- morous responses to a hashtag given by host Chris Hardwick during the program. 2 Background Training Language Models (LMs) is a straight- forward way to collect a set of rules by utilizing the fact that words do not appear in an arbitrary order; we in fact can gain useful information about a word by knowing the company it keeps (Firth, 1968). A statistical language model estimates the probability of a sequence of words or an upcom- ing word. An N-gram is a contiguous sequence of N words: a unigram is a single word, a bigram is a two-word sequence, and a trigram is a three-word sequence. For example, in the tweet tears in Ramen #SingleLifeIn3Words “tears”, “in”, “Ramen” and “#Sin- gleLifeIn3Words” are unigrams; “tears in”, “in Ramen” and “Ramen #SingleLifeIn3Words” are bigrams and “tears in Ramen” and “in Ramen #SingleLifeIn3Words” are trigrams. An N-gram model can predict the next word from a sequence of N-1 previous words. A tri- gram Language Model (LM) predicts the condi- tional probability of the next word using the fol- lowing approximation: P(wn|wn−1 1 ) ≈ P(wn|wn−2, wn−1) (1) The assumption that the probability of a word depends only on a small number of previous words is called …

Content truncated to 500 words