This paper studies joke detection in short text, focusing only on jokes triggered by lexical ambiguity. Following Attardo and Raskin, we treat these jokes as cases where humor arises from a script opposition activated through a logical mechanism such as homography or homophony. Our framework combines contextuals emantic analysis for homographs with phoneme-level similarity for homophones and near-homophones, using CMUdict, weighted Levenshtein distance, and prompt-based reasoning to recover ambiguities that are not visible in spelling alone. Results show that explicit phonetic modeling improves detection of sound-based puns.