REBench: Microbenchmarking Framework for Relation Extraction Systems.

In recent years, several relation extractions (RE) models have been developed to extract knowledge from natural language texts. Accordingly, several benchmark datasets have been proposed to evaluate these models. These RE datasets consisted of natural language sentences with a fixed number of relations from a particular domain. Albeit useful for general-purpose RE benchmarking, they do not allow the generation of customized microbenchmarks according to user-specified criteria for a specific use case. Microbenchmarks are key to testing the individual functionalities of a system and hence pinpoint component-based insights. This article proposes REBench, a framework for microbenchmarking RE systems, which can select customized relation samples from existing RE datasets from diverse domains. The framework is flexible enough to choose relation samples of different sizes and according to the user-defined criteria on essential features to be considered for RE benchmarking. We used various clustering algorithms to generate microbenchmarks. We evaluated the state-of-the-art RE systems using different RE benchmarking samples. The evaluation results show that specialized microbenchmarking is crucial for identifying the limitations of various RE models and their components.