Title: Improving Centering-based Neural Coherence Modeling
Speaker: Sungho Jeon (HITS)
Abstract
Unlike previous work of coherence modeling, Jeon and Strube (2020) introduce a coherence model that takes structural information into account. This model does not rely on human annotations of structural information but combines Centering theory with modern NLP techniques, such as a pretrained language model. It first captures the focus of each sentence, then tracks the changes of focus to construct the structural relationship between groups of coherent sentences. They evaluate their model on two downstream tasks, and it shows state-of-the-art performance.
However, this work brings more questions. While this work considers all items to capture the focus of a sentence, Centering theory only considers noun phrases. Can we make this model better with noun phrases? Then can we propose a novel local coherence model with focus captured on noun phrases? Then can we combine the local and the structural coherence model? With this unified model, can we show when local or structural factors influence more? In this talk, I share our progress and plans to investigate these questions.
