Decomposing Relational Triple Extraction with Large Language Models for Better Generalization on Unseen Data

Despite the significant achievements of existing methods for relational triple extraction (RTE), most of them only have weak generalization for the data unseen or partially seen in the training set. Recently, large language models (LLMs) have attracted more interests for their powers in various natural language processing tasks, but they still suffer from some issues, including less sensitivity to the order of subjects and objects and unsatisfactory output formats. To harness LLMs to achieve more accurate RTE especially on the unseen data, we propose a novel framework LRTE in which the RTE task is decomposed into three sub-tasks of a pipeline: relation extraction, entity extraction, and triple filtering. To evaluate all models’ RTE performance more truthfully, we also refined two RTE benchmarks through removing noisy triples and complementing the missing triples. Our extensive experiments upon the refined datasets demonstrate our framework’s superior performance over the previous competitors.