Temporal relation extraction traditionally employs discrete categories like BEFORE, AFTER, and OVERLAP to characterize how events relate in time. This project develops a continuous-valued approach that captures fine-grained temporal distinctions and event durations using real-valued scales.
The annotation framework decomposes temporal reasoning into two complementary tasks: relation strength and duration assessment. For temporal relations, annotators use slider interfaces to indicate how much earlier one event occurred relative to another, producing scalar values that capture degrees of temporal precedence. For duration, annotators estimate event length on a log-scale from seconds to decades.
The UDS-Time dataset contains annotations for all pairwise temporal relations between events in each document, making it the largest temporal relation dataset to date. The continuous annotations reveal that temporal relations form a spectrum rather than discrete categories, with many event pairs exhibiting partial overlap or vague temporal boundaries.
Neural models trained on this data learn representations that capture both fine-grained temporal distinctions and traditional categorical relations. The continuous framework enables applications including timeline construction, where the scalar annotations can be used to position events along a temporal axis with appropriate uncertainty bounds.
Publications
-
Vashishtha, Siddharth, Benjamin Van Durme, and Aaron Steven White. 2019. Fine-Grained Temporal Relation Extraction. In Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 2906–2919. Florence, Italy: Association for Computational Linguistics.
-
Vashishtha, Siddharth, Adam Poliak, Yash Kumar Lal, Benjamin Van Durme, and Aaron Steven White. 2020. Temporal Reasoning in Natural Language Inference. In Findings of the Association for Computational Linguistics: EMNLP 2020, 4070–4078. Online: Association for Computational Linguistics.