CC BY 4.0Stenzel, Marc RobinMarc RobinStenzelLübbering, MaxMaxLübberingUlusay, BilgeBilgeUlusayUedelhoven, DanielDanielUedelhovenSifa, RafetRafetSifa2023-05-102023-05-102022https://publica.fraunhofer.de/handle/publica/441456https://doi.org/10.24406/publica-133510.24406/publica-13352-s2.0-85148624192Community question answering platforms like Stackoverflow are among the most popular interactive environments on the Internet for individuals to share knowledge. Finding experts to answer questions is one of those platforms' major challenges. To this end, we compare SBERT-Rec and LDA-Rec, two recommender system algorithms which are based on the state-of-the-art transformer architecture and well-established probabilistic topic modeling algorithm Latent Dirichlet Allocation, respectively. Our results show that SBERT-Rec significantly outperforms LDA-Rec in terms of average rank score. While SBERT-Rec excels in an open-world scenario with no presumptions about the underlying subjects of the corpus, LDA-Rec carves out distinct and human interpretable topics inside a niche closed-world corpus. Finally, we provide a novel metric for expert matching evaluation that supports partial experts/non-experts annotations.enBERTCommunity Question AnsweringExpert FindingLatent Dirichlet AllocationRecommender systemsconference paper