Humor Identification from Short Texts Using Roberta and Albert

Abstract

In the recent times most of the chat based applications which are generative in nature suffer from not understanding the humor intent within the text written. In this paper, we showcase a novel approach for classifying text which contains humor using light weight version of BERT called the ALBERT.  Fine tuning the model takes a major step here with the current data as the existing pre trained model is created on Wikipedia corpus. The proposed system uses word embedding mechanism for text and ALBERT Tokenizer for generating the tokens. The approach reduces the training time to 2/3rd and also retains the accuracy at 98.7% where as other Language models like BERT takes more execution time to generate the approximate results. RoBERTa showcases an increase of 0.3% to 0.6% in accuracy as compared to other language models for Humor Classification.