HumorHawk at SemEval-2017 Task 6: Mixing Meaning and Sound for Humor Recognition

Proceedings of the 11th International Workshop on Semantic Evaluations (SemEval-2017), pages 98–102, Vancouver, Canada, August 3 - 4, 2017. c⃝2017 Association for Computational Linguistics HumorHawk at SemEval-2017 Task 6: Mixing Meaning and Sound for Humor Recognition David Donahue, Alexey Romanov, Anna Rumshisky Dept. of Computer Science University of Massachusetts Lowell 198 Riverside St, Lowell, MA 01854 david donahue@student.uml.edu {aromanov, arum}@cs.uml.edu Abstract This paper describes the winning system for SemEval-2017 Task 6: #HashtagWars: Learning a Sense of Humor. Humor de- tection has up until now been predomi- nantly addressed using feature-based ap- proaches. Our system utilizes recurrent deep learning methods with dense embed- dings to predict humorous tweets from the @midnight show #HashtagWars. In or- der to include both meaning and sound in the analysis, GloVe embeddings are com- bined with a novel phonetic representa- tion to serve as input to an LSTM com- ponent. The output is combined with a character-based CNN model, and an XG- Boost component in an ensemble model which achieved 0.675 accuracy in the of- ficial task evaluation. 1 Introduction Computational approaches to how humour is ex- pressed in language have received relatively lim- ited attention up until very recently. With few exceptions, they have used feature-based machine learning techniques (Zhang and Liu, 2014; Radev et al., 2015) drawing on hand-engineered features such as sentence length, the number of nouns, number of adjectives, and tf-idf-based LexRank (Erkan and Radev, 2004). Among the recent proposals, puns have been emphasized as a cru- cial component of humor expression (Jaech et al., 2016). Others have proposed that text is per- ceived as humorous when it deviates in some way from what is expected (Radev et al., 2015). One of the reasons for such dominant position of the feature-based approaches is the fact that the datasets have been relatively small, rendering deep learning methods ineffective. Furthermore, exist- ing humour detection datasets tended to treat hu- mor as a classification task in which text has to be labeled as funny or not funny, with nothing in between, which makes the task considerably simpler. In contrast, the #HashtagWars dataset (Potash et al., 2016b) provided for SemEval-2016 Task 6 assumes that humor can be evaluated on a scale, reflecting the reality that humor is non- binary and some things may be seen as funnier than others. It is also large in size, making it bet- ter suited to the application of deep learning tech- niques. SemEval 2017 Task 6 used the tweets posted by the viewers of the Comedy Central’s @mid- night show, the #HashtagWars segment. Our team participated in subtask A, which was as follows: given a pair of tweets supplied for a given hashtag by the viewers, the goal was to identify the tweet that the show judged to be funnier (Potash et al., 2017). This paper describes the winning submis- sion, and specifically, our systems that took first and second place in the official rankings for the task. Our goal was to create a model that could repre- sent both …

Content truncated to 500 words