@misc{17725, author = {Cise Midoglu}, title = {LLMs and Sports Multimedia: What Can Conversational Agents Offer Soccer?}, abstract = {Large language models (LLMs) have sparked tremendous interest in recent years, with examples such as the GPT, Gemini, LLaMA, and Mistral series receiving widespread public attention, and the applications of LLM based conversational agents (chatbots) rapidly expanding beyond general purpose tasks. While modern chatbots already support multimodality, handling various media modalities such as text, image, audio, and video to an extent, there is room for improvement in terms of contextual understanding and domain specificity. In this talk, we discuss how association football (soccer), one of the biggest sources of sports multimedia content on the Internet, can benefit from the development of custom generative AI chatbots. We describe our experiences with various data curation, fine tuning, prompt engineering, and retrieval-augmented generation strategies, for the purpose of developing a "Soccer Chatbot" able to interact and converse in natural language with different target audiences such as fans, trainers, athletes, referees, scouts, etc., and undertake tasks such as automatic gameplay analysis, summarization, and highlight search and retrieval, based on an enhanced multimodal understanding of soccer games. We provide proof-of-concept implementations and disseminate lessons learned.}, year = {2024}, journal = {Streaming Tech Sweden (STSWE24)}, url = {https://streamingtech.se/}, }