Binary Classifier for Experimental Search of Triggers in Jokes in English

This paper describes the development of a binary classifier differentiating humorous and non-humorous texts. The proposed seq2seq model consists of a pre-trained BERT embedding layer and a Bi-LSTM layer used for sequence classification. Training and validation corpora include 76,000 jokes and non-jokes with identical vocabulary; this is essential in preventing vocabulary choice from being employed as a distinguishing factor between humor and non-humor. Further, this paper also describes the application of the trained neural network in a series of experiments on linguistic transformations of humorous and non-humorous texts. The purpose of these experiments is to identify the essential parts and words, without which the joke ceases to be humorous. Some interdisciplinary theories of humor specify such expressions as triggers [Attardo S., 1994]. Based on the results of quantitative and qualitative analyses, 78 of the jokes from the validation dataset changed the label to the opposite at least once when the text was transformed. At the same time, 16 of the remaining 22 jokes contain explicit or implicit extralinguistic information. T-test, which measured probabilistic estimates of the original and modified texts for each type of linguistic transformation, revealed (keep tense consistent) the most common types of them: deletion of the punchline, deletion of the setup, deletion of 1 to 3 tokens from the beginning of the text, deletion of 1 to 3 tokens from the middle of the text and deletion of all the nouns.