University of Amsterdam at the CLEF 2024 Joker Track

This paper reports on the University of Amsterdam’s participation in the CLEF 2024 Joker track. Our overall goal is to investigate non-literal use of language, such as in humor and wordplay, that are still challenging current information retrieval and natural language processing technology. Our specific focus is to investigate how an effective wordplay detector can be used for the humorous search results or candidate translations, within the context of the track’s humor retrieval, classification, and translation tasks.
Our main findings are the following. First, standard ranking approaches are effective for retrieving relevant sentences given a query, but a pun classification filter is effective to select humorous results. Second, a BERT encoder based classifier obtains reasonable performance in classifying different aspects of humor, with some distinctions being hard for both models and humans. Third, sequence to sequences machine translation models provide high quality descriptive translation, yet preserving the wordplay across languages remains challenging.
More generally, we revisited the CLEF 2023 Joker Track’s Pun Detection task, and were able to build effective neural pun classifiers. The value of these classifiers was demonstrated as a filter on the results of a standard ranker for the Humor-aware IR task of the CLEF 2024 Joker Track.