Relation guided and attention enhanced multi-head selection for relational facts extraction.

Multi-head selection is a reasonable way of extracting relational facts. Though effective, it ignores the interdependencies of relations and disregards the contextual information. In this paper, we propose a relation guided and attention enhanced approach to address the above challenges. Specifically, we predict the relations existing in the input sentence to guide multi-head selection. This strategy helps to model dependencies of relations. Moreover, we use an attention mechanism to leverage the sentential context. The experimental results demonstrate that our approach significantly outperforms the baselines. [Display omitted] • Multi-head ignores the interdependencies of relations and the contextual information. • A novel RP auxiliary task to guide main table-filling task by pre-predicted relations. • The RP layers can also strengthen the complex interdependencies of relations. • The Token-Pair attention mechanism can leverage the sentential context. [ABSTRACT FROM AUTHOR]