Can AI Recognize a Pun? Exploring Ambiguity with BERT

Can artificial intelligence recognize when one word carries two meanings? In our final Natural Language Processing (NLP) presentation, we explore that question through the word “interest,” focusing on financial return and personal involvement. Our team developed a pun evaluation system using BERT embeddings, which represent language numerically, and cosine similarity, which compares those representations. By comparing target sentences with reference sentences representing each meaning, the system evaluates whether a sentence supports the dual meanings needed for a pun. Using thresholds established from 20 pun examples, we tested the system on 50 sentences. It correctly classified 45 out of 50 sentences, achieving 90% accuracy. It identified approximately 85% of puns and correctly rejected 93% of non-puns. In this presentation, we explain: How the system evaluates two meanings within a sentence. How reference sentences and similarity thresholds influence classification. How a conversational agent can help users explore pun interpretations. Why a single-word focus and manually selected references limit broader use. How adding WordNet’s structured word relationships could improve explanations. Our key takeaway: Strong classification results do not mean AI fully understands humor. Human judgment remains essential for selecting meaningful references, interpreting ambiguity, and assessing whether a pun actually works. Join us as we explore what this project reveals about language, humor, and the challenges of evaluating AI. #NaturalLanguageProcessing #ArtificialIntelligence #BERT #ConversationalAI #ResponsibleAI