Tanishq Goel  supervised by Dr. Radhika Mamidi  received his  Master of Science –  Dual Degree in Computational Linguistics (CLD). Here’s a summary of his research work on Beyond the Punchline: Humor Classification and the Balance Between Wit and Offense:

This thesis advances the computational understanding of humor by building on the success of ensemble transformer models for Spanish humor classification, as demonstrated in our prior work at HAHA@IberLEF2021. In that study, we achieved state-of-the-art results (F1: 0.8850) in binary humor detection through ensembles of multilingual BERT, BETO, and auxiliary classifiers, while also securing competitive rankings in regression and multi-label tasks. However, humor’s subjective nature—often teetering between wit and offense—poses ethical risks for AI systems, as evidenced by high-profile failures like Microsoft’s Tay and Meta’s BlenderBot, which inadvertently generated harmful content. Here, we bridge technical innovation with ethical scrutiny. First, we extend our ensemble methodology, analyzing how transformer architectures capture linguistic nuances (e.g., irony, wordplay) in Spanish tweets while highlighting cultural biases in training data that may conflate humor with insults. We then propose strategies to mitigate such risks, including hybrid moderation systems that combine ensemble predictions with rule-based filters, and stress-test our models on edge cases where humor overlaps with offensive language. By linking robust classification to ethical safeguards, this work not only refines humor detection accuracy but also advocates for AI systems that balance creativity with cultural sensitivity. Ultimately, we argue that computational humor analysis must evolve beyond performance metrics to address the societal implications of misclassification, ensuring AI respects the delicate line between laughter and harm. 

July 2025