Event type discovery in open domains aims to automate the induction of event types from completely unlabeled text data. Conventionally, small models utilize clustering techniques to address this task. Nonetheless, the fully unsupervised nature of these methods results in suboptimal performance by small models in this context. Recently, large language models (LLMs) have demonstrated excellent capabilities in contextual understanding, providing additional relevant information for specific task scenarios, albeit with challenges in precision and cost effectiveness. In this paper, we use LLM to guide the clustering of event texts and distill this process into a fine-tuning task for training smaller pre-trained language models. This approach enables effective event type discovery even in scenarios lacking annotated data. The study unfolds in three stages: in action acquisition, leveraging LLMs to extract type-relevant information from each event text, ensuring that the event representations are particular to task-specific details; in clustering refinement and dual-fine-tune, LLMs refine results from both task-agnostic and task-specific perspectives, with the refinement process designed as fine-tuning tasks under different viewpoints to adjust encoders; and in type generation, post-clustering, LLMs generate meaningful event type labels for each cluster. Experiments show that our method outperforms current state-of-the-art approaches and excels in event type discovery tasks even in completely open-domain with no labeled data.

Zero-shot event detection aims to involve the automatic discovery and classification of new events within unstructured text. Current zero-shot event detection methods have not considered addressing the problem more effectively from the perspective of improving event representations. In this paper, we propose dual-contrastive prompting (COPE) model for learning event representations to address zero-shot event detection, which leverages prompts to assist in generating event embeddings using a pretrained language model, and employs a contrastive fusion approach to capture complex interaction information between trigger representations and sentence embeddings to obtain enhanced event representations. Firstly, we introduce a sample generator to create ordered contrastive sample sequences with varying degrees of similarity for each event instance, aiding the model in better distinguishing different types of events. Secondly, we design two distinct prompts to obtain trigger representations and event sentence embeddings separately. Thirdly, we employ a contrastive fusion module, where trigger representations and event sentence embeddings interactively fuse in vector space to generate the final event representations. Experiments show that our model is more effective than the most advanced methods.