An anchor-guided sequence labeling model for event detection in both data-abundant and data-scarce scenarios.

Event Detection plays an important role in automatically extracting events from massive unstructured web data. However, traditional methods often suffer from maladaptation when implemented in varying data contexts: 1) Although "prompt"-based models demonstrate their worth in data-deficient situations, they underperform in data-abundant settings. 2) Contrastingly, "fine-tuning"-based techniques, which outperform "prompt"-based models in the data-abundant scenario, typically fall short in low-resource environments. In response to these challenges, we propose an anchor-guided sequence labeling model that ingeniously merges the aforementioned paradigms to bolster effective event detection in both data-rich and data-scarce situations. Firstly, we formalize event detection as locating triggers in type-aware inputs, eliminating formal discrepancies between the two paradigms and enabling traditional sequence labeling algorithms to detect events under low-data and zero-shot settings. Secondly, this study introduces an innovative anchor-guided attention mechanism that sets up an information exchange conduit between the two paradigms. Lastly, to enhance the utility of anchors, auxiliary outputs are generated by predicting masked tokens in type-aware inputs. Empirical experiments attest to the efficacy of our proposed method in both data-abundant and data-scarce scenarios.