Retrieval-Enhanced Template Generation for Template Extraction

Renyu Wang, Wei Xiang, Zhenhua Wang, Bang Wang

Abstract: Template extraction tasks, such as role-filler entity extraction (REE) and template filling (TF), are classic problems in information extraction. Previous works usually simplify the TF task and focus only on the REE task. Few works attempt to tackle the more challenging template filling task. However, they design sophisticated models for individual REE or TF that can hardly tackle those two tasks concurrently. In this work, we formulate those two tasks as a template sequence generation problem, which can be solved by a unified generation framework. Specifically, we leverage the template to guide the model to extract event entities from a document and fill them into the predefined template. Furthermore, to enhance the model’s understanding of the input document and capture the dependency between similar templates, we employ a retrieval-enhanced approach. The most semantically similar template is retrieved from the training data and augments the current context with similar structures and semantic information captured in the retrieved demonstration. Extensive experiments on the MUC-4 dataset demonstrate the generality and effectiveness of our proposed model. (The source code is available at https://github.com/luwry/RTG4TE).