Many goal-oriented spoken dialog systems lack a social component like small talk which often features in human-human communication. In this work we aim to alleviate part of this problem by generating sentences which have the goal to appeal to the user and increase the probability of a response. Such sentences are suitable to break the ice in the beginning of a conversation and are therefore referred to as "ice-breaking" throughout this paper. Furthermore, we use data from the Twitter account of the user in order to infer the users interests. By generating sentences about these interests we utilize the existence of homophily in social networks. A user study shows that the described system outperforms one which chooses interests at random. Furthermore, we note that 70% of the study participants would answer the system and continue talking on the same topic which was introduced by the generated sentence.
展开▼
