Fig 1.
Story endings generated by different methods.
Words in red are keywords of the story, which are linked to ConceptNet to infer related words (orange nodes) in method (a). Method (b) directly generates results according to the story context. Method (c) reasons the causal events behind each sentence by various relations on the commonsense events knowledge base and incorporates them into the ending generation.
Table 1.
Relations defined in GLUCOSE.
Fig 2.
The model architecture consists of three components, a shared encoder, an inference decoder, and a generation decoder. The causal events inference task uses the shared encoder and the inference decoder to generate causal events for the story context. The story ending generation task applies the shared encoder and the generation decoder, which integrates the causal events and the story context to finish the ending.
Fig 3.
The first mask is the encoder mask, which is bidirectional, and each token accesses the information of other tokens. The second mask is the mask of the self attention in the decoder, which is a left-to-right mask, and each token only accesses information of history tokens. The third mask is the mask of the cross attention in the decoder, where each token access all information from the encoder.
Table 2.
Data split.
Table 3.
The results on the ROCStories.
Table 4.
Human evaluation.
Table 5.
The effection of different relations.
Table 6.
The effect of joint training.
Table 7.
Case study on the test set of ROCStorise.
Table 8.
Case study on the test set of ROCStorise.