Zero-shot Event Detection Based On Ordered Contrastive Learning

Event detection is one of the classic tasks in natural language processing,and it is also a key step in event extraction and knowledge graph construction.It's a traditional solution for event detection to use a large number of labeled samples to train a language model based on deep learning technology,and then classify unstructured text according to predefined event types.However,as the unstructured text contents keep growing on the Internet,many undiscovered event types emerge in an endless stream.The annotating samples for these unknown events will consume a lot of human and material resources.To this end,the zero-shot event detection task is proposed,where the model needs to automatically discover and classify new event types without relying on any annotated samples.The difficulty of zero-sample event detection is mainly reflected in two aspects:the lack of labeled data and type definition of unknown events.The lack of labeled data means that traditional supervised learning models cannot be trained for unknown events,resulting in a greatly reduced generalization ability of the model on unknown types.The lack of type definition prevents the model from identifying trigger words through semantic similarity,which further hinders the model from learning more accurate event type features.To address these challenges,this paper proposes a zero-shot event detection technique based on ordered contrastive learning.The main contributions are as follows:First,this paper proposes an event feature extraction algorithm based on ordered contrastive learning to solve the problem of lack of labeled data.It utilizes a variety of data augmentation methods to construct several comparative samples.According to the partial order relationship of semantic and type similarity between these contrast samples and the original samples,an ordered contrast loss function is designed.By evaluating the distance of the samples in the latent space of the language model,the ordered contrastive learning loss function can help the language model learns the common features between samples within the same event type and distinguishes samples of different event types at the same time.Second,this paper designs a semi-supervised event type prediction algorithm based on prototype network to make full use of the annotation information of known events.In addition to the new event types without sample annotation,some samples of known event types are already annotated in the practical application.The prototype networkbased semi-supervised algorithm defines different types of prototypes in the latent space and calculates different loss functions for samples of known and unknown types,respectively.Being able to take full advantage of supervised signals of known event types,the network helps pre-trained language models adapt to the downstream task of event detection.Thirdly,this paper introduces the prompt-based prediction technology into the zero-shot event detection model to solve the problem of lack of type definition.With a general template with mask token,the trigger word prediction task is converted into the word prediction task of the pre-trained language model.Thereby model doesn't rely on any predefined event types,external part-of-speech analysis tools,and trigger vocabularies.The cumulative error of model and manual participation is greatly reduced.The above three algorithms as a whole system,under the unified problem definition and framework,progressively solve the two significant difficulties of zero-shot event detection and gradually improve the model effect.Finally,this paper performs experiments and proves that the proposed method can achieve better results than existing methods on two English event detection datasets,ACE2005 and Few Shot ED.While ensuring the recognition ability of known types,the F1-score of unknown types is increased by 30.8% and 33.2% respectively.