Using Reverse Lookup and Bidirectional Query for Event Argument Extraction Prompt Enhancement

The rapid growth in China's manned space science and application field has generated numerous field events, offering valuable insights for supporting related research in space science and application domains. The application of prompt tuning technology in space science and applications still encounters challenges in constructing suitable prompt templates for domain events. To address this issue, we introduced a prompt enhancement model, focusing on improvements in three key parts: (1) Reverse lookup module use parameters in the text to find the position of the query in the prompt statement to enhance the prompt information. (2) Bidirectional query use soft prompts to design a prompt statement in the opposite order of the original prompt query to jointly perform the extraction work. (3) Query matrices replace the prompt slots in the previous method, the key idea is to use a trainable matrix to expand a single query into multiple queries for multi-argument extraction. We conducted experiments on four datasets. The results underscore the effectiveness of our method, demonstrating average F1 gains of 3.5% and 2.8% for Bart-base and Bart-large on the four datasets, respectively. Additionally, ablation experiments provide compelling evidence that our improvements are both necessary and effective.