Retrieval-Enhanced Event Temporal Relation Extraction by Prompt Tuning

Rong Luo, Po Hu

Abstract: Event temporal relation extraction aims to automatically identify the temporal order between a pair of events, which is an essential step towards event-oriented natural language understanding and generation. For this task, impressive improvements have been made in neural network-based approaches. However, they typically treat it as a supervised classification task and inevitably suffer from under-annotated data and label imbalance problems. In this paper, we propose a new retrieval-enhanced event temporal relation extraction model, called PRetrieval, which addresses these issues by making full use of the golden relation labels to learn potential temporal knowledge in a pre-trained language model and to perform prompt tuning. Specifically, PRetrieval first generates an initial prediction using the prompt-based input and constructs a datastore using this prediction result and the associated golden labels, which has the advantage of not relying on any external datasets or additional knowledge enhancement. PRetrieval then retrieves examples with outputs similar to the preliminary prediction results as the new prompt to improve the model’s relation extraction ability under few-shot settings. In addition, it also retrieves examples with similar input to further enhance the model performance. The experimental results on two benchmark datasets (i.e., MATRES and TB-Dense) show that our proposed approach significantly outperforms the state-of-the-art methods.