KDEhumor at SemEval-2020 Task 7: A Neural Network Model for Detecting Funniness in Dataset Humicroedit

Proceedings of the 14th International Workshop on Semantic Evaluation, pages 852–857 Barcelona, Spain (Online), December 12, 2020. 852 KdeHumor at SemEval-2020 Task 7: A Neural Network Model for Detecting Funniness in Dataset Humicroedit Rida Miraj and Masaki Aono Department of Computer Science Toyohashi University of Technology, Japan ridamiraj974@gmail.com, aono@tut.jp Abstract This paper describes our contribution to SemEval-2020 Task 7: Assessing Humor in Edited News Headlines. Here we present a method based on a deep neural network. In recent years, quite some attention has been devoted to humor production and perception. Our team KdeHumor employs recurrent neural network models including Bi-Directional LSTMs (BiLSTMs). Moreover, we utilize the state-of-the-art pre-trained sentence embedding techniques. We analyze the performance of our method and demonstrate the contribution of each component of our architecture. 1 Introduction Humor, a way of giving entertainment and provoking laughter, is an exceptionally savvy communicative activity. In recent years, quite some attention has been devoted to humor production and perception. What makes us laugh when reading a funny sentence? Mostly it depends on the cultural area background of a person. Humor cannot be defined in standards as it varies with nation to nation, area to area, people to people. Many types of humor require substantial external knowledge such as irony, wordplay, metaphor and sarcasm. These factors make the task of humor recognition difficult. Recently, with the advance of deep learning that allows end-to-end training with big data without human intervention of feature selection, humor recognition becomes promising. However detecting humor in news headlines from which one word is chosen to be replaced by another word to make the micro-edited sentences look funny is quite challenging. To address the challenges of humor detection in edited news headlines, (Hossain et al., 2020) proposed detection of funniness level in dataset named as Humicroedit1, Task 7 at SemEval-2020. They focus on two related subtasks. Task 1 defines regression problem where a system needs to predict how much funny is the edited news headlines (humicroedit) from range 0 to 3. Whereas Task 2 defines the classification problem where a system need to predict the funniest sentence between two edited headlines. In this paper, we only focus on the Humicroedit data for both Tasks 1 and 2. No extra data, or other sources is used to get more data. The rest of the paper is structured as follows: Section 2 provides a brief overview of prior research. In Section 3, we introduce our proposed neural network model. Section 4 includes experiments and evaluations as well as the analysis of our proposed method. Some concluded remarks and future directions of our work are described in Section 5. 2 Related Research Humor is omnipresent, universal, elusive event which exists in all societies and cultures and fulfills a range of social, cognitive and emotional functions. The task of automatic humor recognition refers to deciding whether a given sentence expresses a certain degree of humor. In previous researches and studies, mostly the problems related to humor were based on binary classification, …

Content truncated to 500 words