Timeline Deliberation for Fine-grained Temporal Ordering

Anonymous

Abstract: Despite recent advances, language models still struggle to capture temporal orders between events. For example, it is not trivial to teach the fine-grained difference between two questions “happened right before” or “happened often before”. Previous solutions have relied on weak supervision, namely answer overlaps, as a proxy label to contrast similar and dissimilar pairs. In contrast, we claim that answer overlap on the question pair is too weak a signal for contrastive learning (also known as shortcut problem). So we propose to leverage question “bundles”, a related question subset we group with respect to the events in the passage, as a stronger supervision to approximate a timeline of a passage. We introduce the Timeline Deliberation Network (TDN), which reasons over the timeline in a two-level process: The drafting layer drafts answers based on semantic and syntactic evidence. The refinement layer aggregates over contrast question groups as a set of inputs and collectively refines answers to maintain temporal consistency. Results on TORQUE and TB-dense datasets demonstrate that TDN outperforms previous methods, by effectively resolving the shortcut problem.

Paper Type: long

Research Area: Semantics: Sentence-level Semantics, Textual Inference and Other areas