Time Expression Normalization with Meta Time Information
Mengyu An, Chenyu Jin, Xiaoshi Zhong, Erik Cambria · 2023
Time expression (a.k.a., timex) normalization is a fundamental task for many downstream researches and applications. Previous researches mainly developed deterministic rules and machine-learning methods for the end-to-end task of timex recognition and normalization (TERN). However, deterministic rules heavily depend on specific domains while machine-learning methods are somewhat unexplainable. To better understand the task, we analyze three diverse benchmark datasets for the characteristics of timex types and values. According to these characteristics, we propose a rule-based method termed MetaTime1with three kinds of meta time information to normalize timexes into standard type and value formats. MetaTime is independent of specific domains and textual types. Experimental results on three diverse benchmark datasets demonstrate that MetaTime outperforms four representative state-of-the-art methods.