A dataset of real dialogues for conversational recommender systems

Conversational Recommender Systems (CoRS) that use natural language to interact with users usually need to be trained on large quantities of text data. Since the utterances used during the interaction with a CoRS may be different depending on the domain of the items, the system should also be trained separately for each domain. So far, there are no publicly available datasets based on real dialogues for training the components of a CoRS. In this paper, we propose three datasets that are useful for training a CoRS in the movie, book, and music domains. These datasets have been collected during a user study for evaluating a CoRS. They can be used to train several components, such as the Intent Recognizer, Entity Recognizer, and Sentiment Recognizer.