Prompt Selection

The prompt is selected using the same technique as is used for the topic classification. For each prompt we pre-compute a model which keeps the word frequencies within the prompt. After the current topic of the conversation is established we find the prompt within the topic which is the most related to the recent utterances as given by the conversation model ${\bf x}$ (Eq. 2). We compute the relevance measure (Eq. 3) with class c being the corresponding prompt model.