Browsing by Subject "Conditional generation"

Now showing 1 - 1 of 1

Conditional generation of temporally-ordered event sequences
(2020-12-04) Lin, Shih-Ting; Durrett, Greg
Narrative schema knowledge is useful for a wide range of tasks related to events. In this work, we tackle the problem of making inferences from partial, unordered sequences of events, which arise from having incomplete knowledge about some underlying scenario. We tackle the problems of temporal event ordering and event infilling, predicting new events which fit into a sequence of existing ones according to temporal and coherence criteria. We unify these problems in a single model, a BART-based conditional generation model learned as a denoising autoencoder. This model operates over sequences of events represented as predicate-argument structures. At training, we take event sequences, shuffle them, delete events, and then attempt to recover the original events to encourage our model to capture more general temporal event knowledge. Our evaluation demonstrates that our BART-based models significantly outperform both a BERT-based pairwise model and a BERT-based pointer network on temporal event ordering. A human evaluation also shows that our models are able to generate events that fit more temporally into the input events compared to GPT-2 models.