Title: Revisiting Graph Pattern Based Text Coherence Modeling with Corpus Level GCN
Speaker: Wei Liu (HITS)
Abstract
Inspired by the text organization theory, recent works considered applying graph patterns into text coherence modelling.
However, those approaches focus mainly on a single document, ignore the underlying correlation between documents. In
this work, we propose a corpus-level GCN based method that is capable of capturing similar graph patterns between
documents. Specifically, we first build a sentence-level graph for each document in the corpus, from where we search for
different subgraph patterns. Then, we construct a heterogeneous graph for the whole corpus, which contains document
nodes and subgraph nodes. The GCN is applied on the heterogeneous graph to model the connectivity relationships.
Experiments on the benchmark dataset show the effectiveness of underlying correlations between documents.
