Browsing by Subject "Graph convolutional networks"

Now showing 1 - 1 of 1

Learning coherent narratives from text-based knowledge graphs
(2020-09-14) Tomkovich, Alexander Richard; Erk, Katrin
Although efforts to reason over large-scale knowledge graphs have continued to gain traction over the last decade, document-level knowledge graphs still remain underexplored. In this work, we build on previous efforts in narrative cloze and one-class clustering to develop a system which takes in as input mixtures of text-based knowledge graphs that meet at shared entity/event nodes and aims to expand subgraphs which correspond to coherent narratives within those mixtures. We develop two main types of synthetic data with two associated tasks: "graph salads" are mixtures of text-based KGs generated from distinct documents which require our inference models to iteratively add a sequence of graph pieces to a starting narrative, while "cloze-style" instances are mixtures of text-based KGs from distinct documents which require the admission/rejection of a single graph piece. We develop corresponding neural models which leverage graph convolutional networks and an attention mechanism to perform inference over these graphs. We show that our data is learnable, and offer suggestions for future improvement of both our data generation procedures and model architectures.