Open domain dialogue response generation: models and evaluation metrics
By Alice Oh
Abstract
Dialogue response generation is an important AI research topic. We are investigating this topic with the open domain perspective, meaning the topics of the dialogues are not pre-specified. I describe a new model called Variable Hierarchical User-based Conversation Model which considers the previous dialogues among users, as well as their social network connections. I then describe a new evaluation metric called Speaker Sensitive Response Evaluation Model which correlates better with human judgments. These are described in our recent EMNLP 2019 and ACL 2020 papers.