Evaluating and Improving AI Systems for Pun Understanding

Date:

Guest Research Talk at Technical University of Munich (TUM), CCBE Library, Munich, Germany

Time: 15:00–17:00

Location: CCBE Library

This guest research talk examines how AI systems understand puns and what pun understanding reveals about phonological reasoning, lexical ambiguity, contextual meaning, and common-sense knowledge.

Abstract

The talk presents APUN-Bench, an audio pun benchmark covering pun recognition, pun-word localisation, and meaning inference. Results from audio-language models and cascaded speech-recognition pipelines highlight persistent challenges in fine-grained localisation and implicit meaning inference.

It also introduces PunGraph, a retrieval-enhanced framework that combines phonetic resources, lexical knowledge, and large language models through a structured knowledge graph. Retrieving plausible words and senses constrains the reasoning space and supports more accurate, controllable interpretation of heterographic and homographic puns.

Together, these studies show how structured phonetic and semantic knowledge can improve the reliability of multimodal language-model reasoning.