Lexicon-based Bidirectional Framework for Relational Triple Extraction

Abstract

A lexicon-based bidirectional extraction framework is proposed to solve the unidirectional frame sensitivity problem in relational triple extraction methods. The framework employs two complementary directions to extract entity pairs, effectively alleviating the overdependence on subject extraction results. A shared encoder is utilized to facilitate feature transfer between two directions, ensuring extractions in each direction are mutually reinforced and complementary. The shared structure leads to inconsistent convergence rates in training process, thus a shared-aware learning mechanism is introduced. Model's efficacy is confirmed by experiments demonstrating the adaptability and enhancement capabilities of bidirectional extraction framework and shared-aware mechanism for other labelling-based methods.