Abstract
In order to generate more content-rich responses, as well as avoid monotonous or irrelevant ones, some researchers set off to focus on dialogue generation models in which causal relation is also taken into consideration. However, it is hard to distinguish whether there is causal relation between speaker’s and responder’s utterances from dialogue automatically, which makes collecting such utterance pairs even harder. In this paper, a transfer learning method for dialogue generation training data collection is proposed: the sentence with causal relation feature is learned from web contents such as Wikipedia in the first place and further used to detect utterance pairs containing causal relation from dialogs. Subsequently, the collected utterance pairs can be used for dialogue response generation, which can make agent’s response retain pre-learned causal relation feature in sentence level according to user’s utterance. It is shown that the proposed method yields good performance by using BERT embedding in detecting causal relation from utterance turn in dialogs corpus.
