To laugh or not to laugh – LSTM based humor detection approach

Humor holds the power to turn any mundane conversation into something more enthralling. It is an important feature of personal communication. Sentiment Analysis helps people as well as corporations to comprehend the attitudes of people towards certain words based on how they are strung together. Particularly, Humor Detection facilitates the understanding of underlying relations of words and how various combinations of strings can invoke laughter. In this paper, an approach to identify the presence of humor in sentences, using a model based on Long Short-Term Memory (LSTM), is adopted. LSTM is a specialized version of Recurrent Neural Network (RNN). While conventional RNNs involve cyclic connections for modeling sequenced data, LSTM uses additional units for retaining information for comparatively long periods. A wide-ranging, evenly distributed dataset is used in the implementation of the model. After preprocessing, the data progresses along the embedding layer, LSTM layer and the dense layer and the outcome is finally compiled. Ultimately, the model states whether a given statement is humorous or not. The proposed LSTM model gives an accuracy of 94.62%. As new sources of obtaining humorous content keep emerging through multimedia, new patterns can be discovered. Understanding such patterns through such existing Humor Detection models can promote the development of Automated Humor Generation Systems.