Task generality in relation extraction

Abstract

Relation extraction involves the identification of relations between entities in text. Many distinct tasks in natural language processing, including semantic role labeling, quotation analysis, and event extraction, can be categorized as instances of relation extraction, and share similar structures. However, despite the similarities between these tasks, modeling approaches tend to show little overlap, and model architectures designed for one type of relation extraction task can rarely be applied to others. This situation stands in contrast to other task paradigms common in natural language processing, such as text classification and text generation, wherein existing architectures tend to be highly generalizable to many distinct tasks within their paradigms. This dissertation investigates task generality for relation extraction, that is, the ability or inability of relation extraction model architectures to be successfully applied to diverse relation extraction tasks. To this end, we make a number of concrete contributions: First, we present a formal description language for specifying the properties of different relation extraction tasks, and introduce a software framework for developing model architectures which can automatically account for these properties. By delineating task-specific frontends from task-general backends, this framework enables task-general architectures to be easily adapted to the specifics of particular tasks. Next, we investigate task generality for span extraction, an important subtask of relation extraction. We identify architecture design choices which facilitate task-% generality, and go on to statistically analyze how different types of architectures generalize to different types of tasks, gleaning insights into which task properties, model properties, and interactions therebetween are important for generalization. Finally, we present a method for enforcing regular-language constraints on the outputs of a class of sequence labeling models. We show how constraints can be constructed which capture the specific structures of relation extraction tasks, such that label sequences can be interpreted as relations. Overall, this dissertation works towards making relation extraction more task-general, and we hope our contributions can spur further work in this direction.