Read, Eliminate, and Focus: A reading comprehension paradigm for distant supervised relation extraction

Distant supervision aligns the unstructured text to the knowledge base, thereby enabling automatic machine annotation. Nevertheless, this inevitably introduces a considerable amount of noise. Distant supervised relation extraction models aggregate all sentences sharing the same entity pairs into bags and employ various attention mechanisms to reduce the impact of noisy instances. However, this approach fails to preserve the contextual semantics associated with entity pair information and fails to leverage the strong semantics derived from labels. To address this issue, we explore a novel passage-level reading comprehension paradigm that simulates the human reading comprehension process by (1) eliminating incorrect options (reducing the influence of noisy labels) and (2) focusing on crucial clues (information about entity pairs). Specifically, we propose a RelFoRE model that first selects the top-k possible relations from the label set, and then obtains the most crucial information about the entity pair through bidirectional interactions between the passage, the question, and the options. Extensive experiments are conducted on three widely used datasets, demonstrating significant improvement of the RelFoRE over other competing methods.