Event type induction method and system based on contrastive learning and iterative optimization

Abstract

The application discloses an event type induction method and system based on contrast learning and iterative optimization. The method first constructs a training data set and then removes multiple event type samples in the training data set. Secondly, trigger words are selected from the training data set, event encoding is performed, and an event type perception model is trained. Then, the obtained event type perception model is used to encode an unknown event data set, and clustering is performed, so that each cluster represents an event type. Finally, the obtained event type clustering cluster is guided by a prompt to guide a large language question and answer model to name the event type, and iterative optimization is performed. The system comprises a multiple event sample removal module, a model training module, a clustering module and a naming module. The application effectively removes multiple event samples, improves the unknown event type discovery effect, and solves the problem of poor model training effect caused by the lack of unknown type event labels.

Images