A Study of Bert-based Improvements in Multi-party Conversational Machine Reading Comprehension

Nowadays, dialogue question-and-answer reading comprehension tasks are developing rapidly, and with the increase of roles, multi-party dialogue machine reading comprehension tasks have become one of the current research hotspots. However, the current popular machine reading comprehension models are not very effective on this task. Aiming at this problem, this paper proposes a new method applied to the study of multi-party dialogue machine reading comprehension, expanding on the Bert-based model by performing cross-entropy computation as well as introducing new parameters for parameter fusion, and finally obtaining a new loss function for loss computation. The experiments compare three state-of-the-art models: QANet, R-Net and BiDAF, and three evaluation metrics, namely Exact Match (EM), span match (SM) and utterance match (UM), are selected for evaluation. The results show that the method has more obvious improvement on FriendsQA dataset.