Data-Scarce Event Argument Extraction: A Dynamic Modular Prompt Tuning Model Based on Slot Transfer

Event Argument Extraction (EAE) facilitates comprehension of the text related to an event by extracting and analyzing event arguments. The superiority of previous studies typically rises from the abundance of high-quality event annotations, which is labor-intensive to produce and hard to satisfy by reality. In this paper, we approach data-scarce EAE in both low-resource and few-shot scenarios, which have far-reaching implications for practice. Specifically, we propose a model called Dynamic Modular Prompt Tuning based on Slot-Transfer (DAMPT), which dispenses with any manual effort usually required in existing methods. DAMPT turns to large-scale language models to generate dynamic modular prompts, which are more adaptable than the static ones manually given by experts. Furthermore, DAMPT incorporates a prompt-tuning algorithm called slot-transfer to facilitate event-specific knowledge transfer. An extensive experimental evaluation validates the effectiveness and generalization ability of DAMPT in data-scarce scenarios.