Recent studies have explored the integration of large language models (LLMs) into caregiving robots. The use of LLMs facilitates the generation of human-like natural dialogues and diverse, varied expressions. However, in case of a lack of direct and specific instructions from a user, gaining an accurate understanding of the user's intent is challenging, which may result in miscommunication. Thus, this study proposed a method to achieve robust communication even in the absence of direct instructions by considering the uncertainty in destination prediction probabilities. The proposed system leveraged GPT-4 to generate various Japanese speech content candidates and employed neighborhood components analysis to facilitate dimensionality reduction. Consequently, notable features of the rich linguistic embeddings were preserved while transforming them into manageable data. The system selected appropriate utterances based on the destination prediction probabilities. Furthermore, this study investigated changes in the speech content according to the variations in these probabilities, and the incorporation of uncertainty in communication was validated.