A Weak Self-supervision with Transition-Based Modeling for Reference Resolution

Abstract

The reference resolution is a task to find the link between an entity and its source action in the same recipe. In this study, we introduce a weak self-supervision method with a transition-based model for reference resolution tasks for recipes, where the aim of the task is to make the syntax of the instructions used for reference resolution with self annotation. The results show that our approach to the problem outperforms the previous unsupervised methods with %8 F1. Especially, our models show %82 accuracies of pronoun, and %85 accuracies for null entity resolution.

Tasks

Reproductions