Multi-level Connection Enhanced Representation Learning for Script Event Prediction

What happens next? Combining enhanced multilevel script learning and dual fusion strategies for script event prediction

“…Similarly, Pichotta and Mooney 21 proposed a multiargument event representation $$v({e}_{s},{e}_{o},{e}_{p})$$, where $$v$$ was the predicate verb describing the event, and $${e}_{s},{e}_{o},{e}_{p}$$ were the subject, object, and prepositional object to the verb, respectively. Since this quadruple structure can express a more specific meaning of an event, hence it has been widely adopted by subsequent works 24,25,30–33,41 . We also follow Pichotta and Mooney 21 in our event representation structure.…”

“…Vo et al 28 proposed an unsupervised event network structure for extracting temporal and causal relations of events at the document level. The above methods mainly focused on coarse‐grained connections at either event or chain level, ignoring more fine‐grained connections between the arguments of events, hence Wang et al 32 employed a masked self‐attention mechanism to model the relations between the arguments of events. Furthermore, they employed a directed graph convolutional network (DGCN) model to model the temporal or causal relations between events in the chain.…”

“…Its key is to obtain a discriminative event embedding representation, hence we first need to consider the representation of arguments that make up an event. Here we adopt a masked self‐attention mechanism from Wang et al 32 to model fine‐grained connections among arguments for learning a better argument representation.…”

“…Inspired by the success of MCer, 32 the chain‐level script learning still employs a DGCN 49 to model the temporal and causal relations between events. The updated event embedding representations are utilized to calculate the relatedness score between the event chain and candidate events for SEP.…”

“…Besides the above single models, we also conduct experiments on the multimodel integration following. [30][31][32] Specially, we combine EventComp, PairLSTM, and MCer into a stronger baseline by weighted summation.…”