Conversational AI 29-DISCOURSE AND DIALOGUE PHENOMENA
There has been a long tradition of rule-based approaches to discourse modelling. In current work methods based on deep learning are being used. Coreference in NLP, reviewing datasets, rule-based and machine learning-based algorithms, and identifying issues for further research, such as: the need for standard evaluation metrics; the need to address confused labeling in datasets due to differences in terms used in different theories of reference; and the need to consider the extent to which world knowledge is required to resolve reference and, if so, how it should be incorporated into the processing.
DETECTING, MAINTAINING, AND CHANGING TOPIC
In an open-domain conversation the participants can talk about a wide variety of topics. Engaging effectively in a conversation requires the ability to detect topics raised by the other participant, maintain the topic as required, detect when the topic has changed, and proactively suggest new topics. As well as being able to track topics and generate relevant responses, a participant in a dialogue should also be able to decide whether to keep on the same topic and when to change to another topic. A knowledge graph to enable a dialogue agent to proactively select new topics in a conversation. Using a conversation dataset constructed from the knowledge graph, a dialogue system was trained to move between topics while keeping the conversation natural and engaging.