A Template Retrieval Approach for Generating Diverse Dialogue Responses: Possibilities and Limitations.

The use of templates provides a good control over the outputs generated by a model, that is an obsvious feature in task-oriented dialogue system such (e.g., restaurant reseravation, flights etc., ). However, repeated use of the same templates for multiple turns and users makes the generated utterances identical, monotonous and robotic. Additionaly human conversation often contains person specific miscellaneous information such as booking a restaurant for celebrating wedding. Here the "wedding" is not directly releated related to restaurant booking. Therefore, such diversity in generated responses can similaute the user's diversity behavior and can be used to better train a dialogue systems. Defining or annotating multiple different templates for the same dialgue semantics is expensive and time consuming. In this paper, we investige the positiblies of generating diverse responses from templates that are automatically constructed. From th ereal human utterances available in the publicly available dialogue datasets, we first construct the template pool. We then use a template retriver to find the top-k useful templates and use each of them when generating a reponse uuterance. Our study shows that when the number of in-domain candidate templates is large enough, this appraoch can be an effective diverse reponse generator.