A conversational model for eliciting new chatting topics in open-domain conversation. (December 2021)

• Record Type: Journal Article
• Authors: Li, Weizhao; Ge, Feng; Cai, Yi; Ren, Da
• Is Part Of: Neural networks. Volume 144(2021)
• Page Start: 540
• Page End: 552
• Publication Date: 2021-12
• Subjects: Response generation -- Proactive conversation -- Topic elicitation -- Open-domain; Neural computers -- Periodicals; Neural networks (Computer science) -- Periodicals; Neural networks (Neurobiology) -- Periodicals; Nervous System -- Periodicals; Ordinateurs neuronaux -- Périodiques; Réseaux neuronaux (Informatique) -- Périodiques; Réseaux neuronaux (Neurobiologie) -- Périodiques; Neural computers; Neural networks (Computer science); Neural networks (Neurobiology); Periodicals; 006.32
• DOI: 10.1016/j.neunet.2021.08.021
• Languages: English
• ISSNs: 0893-6080

Abstract

Abstract: In human conversations, the emergence of new topics is a key factor in enabling dialogues to last longer. Additional information brought by new topics can make the conversation more diverse and interesting. Chat-bots also need to be equipped with this ability to proactively elicit new chatting topics. However, previous studies have neglected the elicitation of new topics in open-domain conversations. At the same time, previous works have represented topics with word-level keywords or entities. However, a topic is open to multiple keywords and a keyword can reflect multiple potential topics. To move towards a fine-grained topic representation, we represent topic with topically related words. In this paper, we design a novel model, named CMTE, which focuses not only on coherence with context, but also brings up new chatting topics. In order to extract topic information from conversational utterances, a Topic Fetcher module is designed to fetch semantic-coherent topics with the help of topic model. To equip model with the ability to elicit new topics, a Topic Manager module is designed to associate the new topic with context. Finally, responses are generated by a well-designed fusion decoding mechanism to explicitly distinguish between topic words and general words. Experiment results show that our model is better than state of the art in automatic metrics and manual evaluations.