Time expression normalization based on multi-scale classification and temporal focus model with hierarchical discourse transfer

Time expression normalization has an important role in natural language understanding. Based on time expression recognition, it further annotates temporal semantic attributes. Due to the complexity and diversity of time expression, its normalization is a difficult task. This paper proposes a semiautomatic normalization method, which integrates multi-scale classification of time expression and temporal focus model with discourse structure transfer. Firstly, we hierarchically classify time expressions, separating the time expression recognition, semantic classification and temporal relations classification from normalization task in order to reduce the cost of handmade rules, and provide the operable basis for establishing normalization rules. Secondly, discourse structure transfer model builds the dynamic mechanism for tracking temporal focus. Experiments on ACE 2007 corpus demonstrated the validity of the proposed method, which make a foundation for future work.