Better Semantic Representation: A Low-Shot Relation Extraction Method Based on Token-Generated Contributions

Anonymous

TL;DR: Better Semantic Representation: A Low-Shot Relation Extraction Method Based on Token-Generated Contributions

Abstract: In light of the era of information explosion, traditional relation extraction methods are in a bottleneck due to data limitations in the face of the constant emergence of new relation categories. Therefore the study of low-shot relation extraction in real scenarios is crucial. In the few-shot scenario, it is necessary to build up the model's ability to summarize the semantics of instances. In the zero-shot scenario, it is necessary to establish the label matching ability of the model. Although they need to establish different basic abilities of the model, the common point is that they all need to build excellent semantic representations in the end, which is ignored by the existing methods. In this paper, we propose a method (TGCRE) based on token-generated contribution to unify low-shot relation extraction by generating better semantic representations. Further, we propose a multi-level spatial semantic matching scheme in zero-shot scenarios, in order to solve the problem of the single matching pattern of existing methods. Experimental results show that our method outperforms previous robust baselines and achieves state-of-the-art performance.

Paper Type: long

Research Area: Information Extraction

Contribution Types: Approaches to low-resource settings

Languages Studied: English